Fansher

etal.

Cognitive Research: Principles and Implications (2025) 10:7

https://doi.org/10.1186/s41235-025-00613-w

ORIGINAL ARTICLE

Open Access

permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the

original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or

other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line

to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory

regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this

licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/.

Cognitive Research: Principles

and Implications

Narrative visualizations: Depicting

accumulating risks andincreasing trust indata

Madison Fansher

, Logan Walls

, Chenxu Hao

, Hari Subramonyam

, Aysecan Boduroglu

, Priti Shah

and

Jessica K. Witt

Abstract

In contexts where people lack prior knowledge and risk awareness—such as the COVID-19 pandemic—even truth-

ful visualizations of data can seem surprising. This can lead people to mistrust the veracity of the data and to dis-

count it, leading to poor risk decisions. In this work, we illustrate how narrative visualizations can achieve a balance

between the beneﬁts of three common risk communication mediums (static visualizations, interactive simulations,

and aﬀect-laden anecdotes). We demonstrate empirically that viewing a narrative visualization mitigates the reduced

concern induced by a static visualization when communicating COVID-19 transmission risk (Study 1). Through media-

tion analysis, we show that narrative visualizations are more eﬀective than static visualizations at increasing concern

about large risks because they increase one’s perceived understanding and trust in data (Study 2). We argue that nar-

rative visualizations deserve attention as a distinct class of visualizations that have the potential to be powerful tools

for scientiﬁc communication (especially in contexts where data are surprising, and empiricism is important).

Keywords Data visualization, Misinformation, Risk perception

Introduction

During our lifetimes we repeatedly expose ourselves to

extremely small risks. For example, the risk of experienc

ing skin cancer increases with exposure. to UV radiation.

To date, there has been a dearth of research on how to

best communicate small but accumulating risks, such

as the risks associated with the continuous exposure to

small amounts of foreign/dangerous substances (e.g.,

heavy metal poisoning) or the repetition of risky behav

iors (e.g., driving without a seat belt; Slovic,et al., 1978).

Prior research suggests that cumulative risks tend to be

underestimated on average (Doyle, 1997; Slovic, 2000),

though this average may be based on a signiﬁcant num

ber of individuals who severely underestimate risk and

another who slightly overestimate risk (De La Maza

etal., 2019). ese types of risks are harder to communi

cate not only because humans are typically bad at accu-

mulation-based judgments, but also because the overall

accumulated risk is perceived to be somewhat implausi

ble, leading to amistrust of data. us, communicative

eﬀorts need to focus not only on identifying the most

appropriate format to depict the accumulation pattern,

also onﬁnding ways to increase people’s trust in the end

result.

People commonly use charts or other static visu

alizations to communicate information about complex

systems like climate change and global pandemics (Fran

coneri etal., 2021). For instance, a visualization designer

might use an icon array that illustrates a rise in tempera

ture over time to highlight the eﬀect of human activity

on climate. Such visualizations eﬀectively convey spe

ciﬁc data or outcomes (e.g., that global temperatures

*Correspondence:

Madison Fansher

mfansher@umich.edu

Department of Psychology, University of Michigan, Ann Arbor, USA

Department of Intelligent Systems, Delft University of Technology, Delft,

The Netherlands

Graduate School of Education, Stanford University, Stanford, USA

Department of Psychology, Koç University, Istanbul, Turkey

Colorado State University, Fort Collins, USA

Department of Physical Medicine & Rehabilitation, University

of Michigan, Ann Arbor, USA

Page 2 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

have risen alongside carbon emissions), but they do not

illustrate the underlying mechanisms that generate them

(e.g., the greenhouse eﬀect). As a result, viewers may not

understand the visualization’s intended message (Newell

etal., 2016) and may even discredit the data if it seems

surprising or inconsistent with their prior beliefs (Lord

etal., 1979; Rhodes etal., 2014; Shah etal., 2017). From

an individual’s perspective, mistrust in data could be

related to various personal factors such as one’s political

identity (Peck etal., 2019) and one’s lack of understand

ing of how the data arise—especially in data showing

complex systems such as the exponential growth of dis

ease prevalence (Fansher etal., 2022a; Witt etal., 2022)

or visualizations illustrating uncertainty (Padilla et al.,

2022a). Furthermore, one’s misunderstanding of the pur

pose of science models (e.g., believing that these models

should make accurate predictions or depict reality) may

also lead to distrust in data (Witt etal., 2022). In the cur

rent study, we examined methods for communicating

cumulative risk: namely with narrative visualizations,

static visualizations, and anecdotes. We also examined

whether one’s trust and understanding of the data medi

ated the relationship between the presence of a visualiza-

tion and change in behavior or risk understanding.

The role ofanecdotes

One approach for risk communicators is to avoid dis-

cussing the data altogether, and instead give an anec-

dote about an individual that was negatively aﬀected by

the risk. is approach is often used to increase concern

about low-risk events. In fact, prior research has shown

that anecdotes may be more persuasive than providing

statistical data in health-related contexts, especially when

personal decisions are involved or when threats are seri

ous (for a review/meta-analysis, see Freling etal., 2020).

Anecdotes are thought to be eﬀective because they are

concrete and easier to understand than statistical evi

dence, are emotionally interesting, and, possibly, more

memorable (Freling etal., 2020). For example, success

ful texting and driving interventions often rely on stories

about texting and driving accidents rather than present

ing data about accident risk (Cutello etal., 2020). In gen-

eral, stories that tug at one’s emotions, increase fear, and

provide visual imagery tend to be persuasive (Cutello

etal., 2020).

Unfortunately, communicating risks solely with anec

dotes is potentially problematic and may hinder criti-

cal thinking. For example, Rodriguez etal. (2016) found

that people were less likely to notice ﬂawed conclusions

when anecdotes were presented in descriptions of sci

ence studies in the media. Furthermore, people often

become inured to anecdotes after multiple exposures

(Slovic etal., 2017). Importantly, anecdotes are also not

eﬀective in all situations (Zebregs etal., 2015). A meta-

analysis by Zebregs etal. (2015) found that statistical data

were generally more inﬂuential on attitudes and beliefs

than anecdotes, while anecdotes were more eﬀective at

changing behavioral intentions. For example, anecdotes

were less likely to alter readers’ attitudes toward climate

change, seat belt use, or exercise than statistical data. In

contrast, anecdotes had a larger inﬂuence than statistical

data on behavioral intentions with respect to risky sexual

behavior, exercising, and tanning bed use. A more recent

and somewhat broader meta-analysis suggests that anec

dotes can also be more persuasive than statistical data

when the risk consequences are high, and when they are

related to personal decisions rather than decisions about

others. Interestingly, one study found that individuals

reported preferringstatistical data to anecdotes for mak

ing decisions (Freling etal., 2020). us, although anec-

dotes may be able to increase risk concerns (at least when

risks are severe and personal), they are not a panacea for

communicating risks to the public. ese drawbacks call

for a method to convey data eﬀectively without relying

on anecdotes. One possibility is the use of data visualiza

tions depicting risk.

Visualizing risks

Static visualizations

One approach to help individuals understand risk and

frequency data, especially individuals with low numeracy,

is to present icon arrays or similar visualizations (Galesic

& Garcia-Retamero, 2011). Often, pairs of icon arrays are

used to present relative and absolute risks. Icon arrays

are especially helpful for communicating just how small

risks might be. For example, a recent study showed that

presenting a very small risk (i.e., the risk of serious side

eﬀects from the COVID-19 vaccine) as an icon array

(one in 1 million dots) was eﬀective in communicating

the rarity of a ~ 0.000001% risk (Fansher etal., 2022b).

is representation led to much lower concerns about

COVID-19 vaccines and increased positive attitudes

toward vaccination.

While prior work demonstrates that visualizations like

icon arrays are eﬀective at communicating risks, icon

arrays are still static visualizations, meaning that they

provide no mechanistic explanation for how the larger-

than-expected (or smaller-than-expected) risk calcu

lations are derived—they merely present the ﬁnal risk

estimate. erefore, they may not be as compelling to

the viewer, and may be less likely to increase trust in data

when compared to visualizations that include informa

tion about how the risk estimate was derived.

Page 3 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

Interactive simulations

One way to provide people with a richer understanding

of data is to provide them with interactive simulations

that demonstrate how data change under diﬀerent sce

narios. Interactive simulations change how people view

data, especially data about diﬀerent potential future out

comes. Rather than thinking of outcomes as ﬁxed, they

recognize that science models are tools for examining

the consequences of diﬀerent changes (Herring et al.,

2017; Sterman, 2011). For example, Herring etal. (2017)

gave participants simulations that showed how diﬀerent

emissions scenarios led to diﬀerent magnitudes of cli

mate change. ey found that using those simulations

increased people’s concerns about climate change. Simi

larly, Witt etal. (2022) showed that interacting with sim-

ulations explaining exponential growth of disease spread

as well as the impact of various public health policies

increased people’s understanding and trust in social dis

tancing recommendations.

Although interactive simulations are often eﬀective,

they have several potential limitations. First, interac

tive simulations require more time and resources to cre-

ate than static simulations. Second, individuals must be

motivated to interact with simulations and manipulate

relevant variables. If too little or too much guidance is

provided for exploring a simulation, individuals may

simply not engage (Adams etal., 2015). ird, individu

als can engage in random or mindless interactions rather

than systematically exploring the eﬀects of diﬀerent vari

ables (Liew etal., 2014) and such visualizations may con-

tain elements that are distracting (Hegarty, 2004; Mayer,

2005). Finally, simulations require individuals to under

stand the mapping between diﬀerent visualizations,

which is often challenging (Magana & Silva Coutinho,

2017).

Narrative visualizations

e two types of visualizations mentioned above fall

at the extreme ends of a spectrum. Static visualizations

communicate eﬃciently but oﬀer little mechanistic

explanation, while simulations oﬀer detailed mechanis

tic information but require more of the viewer’s time and

the designer’s careful guidance. In contrast, narrative vis

ualizations (Segel & Heer, 2010) strike a balance between

these two extremes. ese visualizations use step-by-step

explanations to convey details (like simulations) but con

strain the viewer’s inquiry to a predeﬁned set of insights

(like static visualizations). Visualizations that meet this

deﬁnition already appear in popular publications such

as the New York Times (Buckley etal., 2022; Byrd etal.,

2022) and Reuters (Cage, 2021; Dutta etal., 2019; Levine

etal., 2021), but to our knowledge they have not been

identiﬁed as a speciﬁc group of visualizations or studied

speciﬁcally in the context of risk communication. Okan

etal. (2015) examined diﬀerent factors that may improve

the eﬃcacy of icon arrays for communicating risk. ey

found evidence that showing data in the form of icon

arrays alongside explanatory labels increased risk under

standing, particularly in participants with low graph lit-

eracy, suggesting that narrative visualizations may be an

eﬀective tool for presenting risk information.

Narrative visualizations have been widely used by visu

alization designers to ensure readers attend to key data

related facts (Bach etal., 2018; Hullman & Diakopoulos,

2011; Hullman etal., 2013; Segel & Heer, 2010). Typically,

these visualizations incorporate external guidance such

as sequential panels (e.g., the martini glass approach)

and annotations to draw readers’ attention through the

data in the appropriate order. In these approaches, the

goal is to control how readers comprehend the visuali

zation by ensuring coverage of key facts and associated

narrative context, but not necessarily to increase under

standing about the data-generating process (i.e., how

the facts occurred). For instance, Lee etal. (2021) high

lighted the ways in which coronavirus skeptics eﬀectively

used counter-visualizations to promote their anti-mask

stances. Reinholtz etal. (2021) showed that participants

presented with inﬂow visualizations (new coronavirus

cases each day) judged coronavirus risk as lower com

pared to participants presented with stock visualizations

(cumulative number of coronavirus cases). e subset

of narrative visualizations that focus on explaining the

data-generating process might be especially appealing

for complex and empirically driven communication chal

lenges like cumulative risk.

Trust anddata visualization

e literature on trust in data visualization has prolifer-

ated in recent years, likely due to a need for improved

scientiﬁc communication in an era where misinformation

threatens public health (Cook et al., 2015; Roozenbeek

etal., 2020) and people are increasingly skeptical toward

science (Rutjens etal., 2018). While data visualizations

have been shown to be eﬀective at communicating risk

information (Lipkus & Hollands, 1999), especially in the

context of health risks (Garcia-Retamero et al., 2012),

researchers suggest that trust in the presented data is

necessaryin order to update their beliefs and behaviors

accordingly (Garcia-Retamero & Cokely, 2017; Mayr

etal., 2019). Researchers have examined how the features

of visualizations inﬂuence trust in various contexts, such

as viewing election forecast visualizations (Yang et al.,

2023), and the role of visualization features such as pro

cessing ﬂuency (Elhamdadi etal., 2022a), title alignment

(Kong etal., 2019), and the communication of uncertainty

Page 4 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

(Gustafson & Rice, 2020; Hullman, 2019; Kerr etal., 2023;

van der Bles etal. (2020); Yang etal., 2023).

ere are many challenges with measuring and deﬁn

ing trust in the context of data visualization (Elhamdadi

etal., 2022b, 2023). In the current study, we adopt the

general deﬁnition of trust in visualization from Mayr

etal, (2019, p. 25), in that “trust is the user’s implicit or

explicit tendency to rely on a visualization and to build

on the information displayed.” Since the time of this

study, other authors have presented multidimensional

frameworks describing the various facets of trust in visu

alization (e.g., Elhamdadi etal., 2023; Pandey etal., 2023).

For example, Pandey etal. (2023) described trust in visu

alization as being based on data credibility, clarity, relia-

bility, familiarity, and conﬁdence. Trust in visualization is

also commonly referred to as a type of interpersonal trust

between the trusted (visualization) and trustee (user)

(Kelton et al., 2008; Lewicki et al., 2006; Zhao, 2017).

Wang etal. (2014) described trustworthiness as referring

to the honesty and believability of the source, with peo

ple being more likely to adopt suggestions from trustwor-

thy sources and be inﬂuenced toward positive attitude

change.

Indeed, in Garcia-Retamero and Cokely’s (2017) review

on the eﬀect of data visualization on risk literature, the

authors state that “Well-designed visual aids robustly

improve risk understanding by encouraging more thor

ough deliberation, facilitating self-assessment, and

reducing biased risk representations, which in turn ben

eﬁt attitudes, behavioral intentions, and trust, leading to

healthier decisions and more positive health outcomes”

(p. 622). ere is a plethora of research suggesting that

visual aids can lead to improved risk understanding (e.g.,

Garcia-Retamero etal., 2012, Garcia-Retamero & Cokely,

2017; Lipkus & Hollands, 1999; Zipkin et al., 2014),

especially for individuals with low numeracy (Garcia-

Retamero & Cokely, 2014). ome researchers suggest that

data visualization can alsolead to increased trust in data.

Petrova et al. (2015) found that viewing visualizations

of prostate cancer risk was associated with increased

risk comprehension which was in turn associated with a

willingness to engage in shared decision-making with a

physician, suggesting that viewing the risk visualization

increased trust in the provider.

The current study

Anecdotes, static visualizations, interactive simulations,

and narrative visualizations are all used to communicate

data about risk in popular press, but each contain dif

ferent types and amounts of information, as discussed

above. In the current study, we examine the potential

beneﬁts of narrative visualizations for communicating

cumulative risk. Prior research on visual communication

of risk oﬀers guidance on increasing concerns about

underestimated risks, but itdoes not address thecom

munication of cumulative risk. Rather, much of the prior

work focuses on helping individuals comprehend mag

nitudes of individual risks. For example, in public health

contexts perceived risk is a necessary precondition for

people to change their behavior (van der Pligt, 1996).

In this work, we focus on narrative visualizations

because they convey information about data lineage—

which is typically not conveyed by some other commonly

used mediums like static visualizations and anecdotes.

We argue that, because they communicate information

about the data-generating process, narrative visualiza

tions could be powerful tools for scientiﬁc communica-

tion- particularly in risk contexts where people may be

harmed if they make decisions based on inaccurate or

incomplete information. We used the COVID-19 pan

demic to empirically evaluate the eﬀectiveness of nar-

rative visualizations for communicating accumulating,

quantitative risk information.

Study 1 compared the eﬀectiveness of narrative visu

alizations with other common risk visualizations for

communicating the cumulative risk of COVID infection

at a anksgiving dinner. e inﬂuence of communi

cated data on people’s attitudes and decision-making can

also depend on people’s trust in data, calling for more

focused interdisciplinary research on the relationship

between trust and data visualization (Borgo & Edwards,

2020; Mayr etal., 2019; Park & Gil-Garcia, 2022). In a

follow-up study, we investigated the mechanism behind

the narrative visualization’s eﬀectiveness and found that

narrative visualizations increased trust in the data by

increasing feelings of understanding, while anecdotes did

not signiﬁcantly aﬀect understanding.

Study 1

e goal of Study 1 was to use the context of the COVID-

19 pandemic to test the eﬀectiveness of narrative visu

alizations for communicating accumulating, quantitative

risk information. In November 2020, just a few days prior

to the USA anksgiving Holiday, we tested how diﬀerent

data presentation methods (i.e., anecdotes, static visu

alizations, and narrative visualizations), diﬀering in their

amount of information about the data-generating process

and appeal to emotion, would diﬀerentially impact peo

ple’s COVID-19-related plans and concerns during the

anksgiving season. e materials for all the conditions

were designed to closely mimic COVID visualizations

found in the popular press, while prioritizing legibility of

the text and graphs. We tested participants immediately

before anksgiving and in a follow-up study a few weeks

after to conﬁrm whether the presentation in fact led to

Page 5 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

any changes in plans. e complete materials, code, and

data are available at https:// osf. io/ k43ev/? view_ only=

67cc6 401b0 69468 89ac4 99fc1 9afec 2f. For exact word

ing of all the survey items, please see the Supplementary

Materials.

Methods

Participants

Complete data were collected from 1592 US partici-

pants (770 female; age M(SD) = 33.54(11.96)) recruited

from Proliﬁc on Nov 25, 2020, a day before anks

giving, to complete session 1. Participants received an

average pay of $10.30/hr (USD), and the median com

pletion time for the study was 4min and 40s. A total of

1276 participants (622 female; age M(SD) = 34.02(12.10)

years) returned for the second part of this study (ses

sion 2) conducted from December 22, 2020 to January

1, 2021. Participants received an average pay of $9.29/

hr (USD), and the median completion time for the

study was 5min and 10s. All procedures were deter

mined to be exempt by theUniversity of Michigan IRB.

Design

Participants were randomly assigned to one of four

conditions where they viewed either a static visualiza

tion (Static Condition, n = 398), the same static visuali-

zation combined with an anecdote (Static + Anecdote

Condition, n = 399), an anecdote presented without a

visualization (Anecdote Condition, n = 398), or a nar

rative visualization (Narrative Visualization Condition,

n = 397).

Materials

Pre-Intervention Materials. Participants were asked to

indicate who they planned to spend anksgiving with

(alone/with household, extended family, friends/neigh

bors, strangers). To measure concerns about COVID-

19, participants responded to two questions:

1. How concerned are you about getting COVID-19/

Coronavirus at anksgiving?

2. How concerned are you that someone in your family

will get COVID-19 at anksgiving?

Participants rated their level of concern for each of

these questions on a scale of 0 (not concerned at all) to

100 (extremely concerned) using slider scales (anchored

at 50). To measure participants’ perceived risk of din

ing with an infected individual at anksgiving dinner,

we asked: “What is the risk that at least one person at

a anksgiving table with 10 people has COVID-19?”,

which they responded to on a scale of 0–100% with a

slider scale.

Intervention materials Narrative Visualization Condi

tion. Our narrative visualization intervention presented

participants witha series of icon arrays illustrating the

spread of COVID in a ﬁctional community during anks

giving. e narrative visualization provided a step-by-step

explanation for how there is a 40% risk of COVID trans

mission at a anksgiving dinner of 10 people if there was

a 4% infection rate in the ﬁctional community (Fig.1). e

narrative visualization materials were speciﬁcally inspired

by the FiveirtyEight article, “Why Even A Small anks

giving Is Dangerous” (Koerth & Elena, 2020).

Static Visualization Condition. e Static Visualization

Condition served as a baseline to compare withthe other

conditions. is condition was not intended to evoke an

emotional reaction like an anecdote and did not provide

information about how the data were calculated like the

narrative visualization. Participants in this condition

viewed the last page of the narrative visualization (see

Fig.1, page 7)—a bar graph showing that there would is

a 40% risk of transmission at a anksgiving dinner with

10 peoplein this scenario. Participants were not provided

with a step-by-step explanation about how that number

was calculated.

Anecdote Condition. Participants in the Anecdote

Condition read an anecdote describing an individual’s

personal experience with COVID-19 after contracting it

at Canadian anksgiving (which occurs prior to USA

anksgiving). e anecdote was written by conduct

ing an informal survey of social media posts describing

people’s experiences with COVID-19. e details from

multiple posts were merged with the goal of creating

a cohesive story to describe a serious case of COVID-

19, evoking an emotional response (see Fig. 2 for the

anecdote).

Static + Anecdote Condition. e Static + Anecdote

Condition (see Fig.3) presented the same graph as the

static intervention, but included a text box containing

one representative paragraph from the Anecdote Condi

tion’s materials (Fig.2). We included the Anecdote and

Static + Anecdote Conditions to see how narrative visu

alizations compared to typically successful interventions

in the literature on medical risk assessment (i.e., using

emotional-laden content, rather than explaining the data

or showing data without information on how they were

generated) (Fig.4).

Post-Intervention Survey. After viewing the interven

tion materials, participants completed a post-interven-

tion survey which again assessed their attitudes and

concerns regarding COVID-19. ey answered the

same two concern items and the perceived risk item as

in thepre-interventionsurvey. ey were also asked the

question: “How do you feel about your current plans?”

to which they responded with a slider scale from 0 (not

Page 6 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

concerned at all) to 100 (extremely concerned). While

change in concern, change in current plans, and change

in perceived risk are the main variables of interest in our

analysis of theSession 1data, participants also answered

demographic questions, and disclosed their plans for the

December/January holidays. All items are available in the

Supplementary Materials.

Post-anksgiving Follow-up Survey. To assess the

longevity of any changes in concern, and to determine

whether the narrative visualization aﬀected participants’

actual behavior (i.e., holiday plans), we followed up with

participants after anksgiving. For all items, please see

the Supplementary Materials.

Participants were ﬁrst asked whether they planned to

attend or host/have attended or hosted any in-person

gatherings with people outside of their household for

the December/January holidays (Yes, No, Unsure). We

also asked whether COVID had changed their holiday

plans (No change; spending the holidays with fewer peo

ple than planned; spending the holidays with more peo-

ple than planned). We again assessed perceived risk of

COVID with the item: “Based on the previous Survey of

Holiday Plans, what is the risk that at least one person at

Fig. 1 Narrative visualization intervention materials for Study 1. The page numbers were added to the ﬁgure to show the order in which

the materials were presented to participants

Page 7 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

a anksgiving table with 10 people might have COVID-

19?” to which participants responded with a slider scale

from 0 to 100%.

To see whether intervention condition inﬂuenced par

ticipants’ anksgiving activities, we asked them “Did

you make any last-minute changes to your anksgiving

plans?” (yes/no) followed by “Did your change of plans

increase or decrease the number of people you interacted

with during anksgiving?” (increased, decreased, nei

ther increased nor decreased). We also assessed COVID

concern with two items:

1. How concerned are you about getting COVID-

19/Coronavirus via social gatherings in Decem

ber? (slider scale of 0 (not concerned at all) to 100

(extremely concerned))

2. How concerned are you that someone in your family

will get COVID-19 via social gatherings in Decem

Fig. 2 Anecdote intervention materials. Note. Participants read an anecdote from someone who had contracted COVID at Canadian Thanksgiving.

The below is a representative excerpt from the anecdote. The full text is one page in length and is available in the Supplementary Materials

Fig. 3 Static + Anecdote Condition materials for Study 1. Note. Participants in the Static Condition viewed the bar graph without the anecdote

Fig. 4 Study 1 task ﬂow

Page 8 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

ber? (slider scale of 0 (not concerned at all) to 100

(extremely concerned))

Procedure

Proliﬁc workers were invited to participate in our ini-

tial survey, conducted before American anksgiving

on November 25, 2020. After providing informed con

sent, participants completed the pre-intervention sur-

vey questions and were randomly assigned to view the

materials associated with one of the four experimental

conditions (Static, Static + Anecdote, Narrative Visu

alization, or Anecdote). ey then completed the post-

intervention survey questions followed by Subjective

Numeracy Scale (SNS-3, McNaughton etal., 2015) and

demographic questions. A few weeks later, participants

from the ﬁrst part of the study were invited to complete

the Post-anksgiving survey. is survey was available

to participants from December 22, 2020 to January 1,

2021. Participants again provided informed consent and

answered the questions described above. e exact word

ing and ordering of all the questions from Study 1 are

available in the Supplementary Materials.

Results

e results presented are guided by our main questions

of interest:

1. Are narrative visualizations more eﬀective than static

graphs at increasing concern and perceived risk?

2. How eﬀective are narrative visualizations relative to

anecdotes?

3. Does adding data to the anecdote condition

(Static + Anecdote Condition) make the intervention

more eﬀective than just viewing one or the other?

Given that we present the results of 20 various tests, we

used a Bonferroni correction to reduce the likelihood of

Type I error. us, in the following analysis, we decreased

our signiﬁcance level to p < 0.003 (

.05

0.0025

Data exclusion

Our analysis focused on participants who indicated

before intervention that they planned to have a meal with

people outside of their home (n = 532), as these were the

participants who were not taking proper precautions. It

can beassumed that that they perceived a lower risk of

contracting COVID-19 than was appropriate. Please see

Table1 for the demographic characteristics and sample

sizes for the participants from each session after applying

our exclusion criteria.

Pre‑thanksgiving data analysis

Change in Concern. We calculated pre- and post-inter-

vention concern scores for each participant by averaging

their responses to the two concern questions in each of

the pre- and post-intervention surveys (i.e., How con

cerned are you about getting COVID at anksgiving?

How concerned are you that someone in your family will

get COVID at anksgiving?). Note that the responses

analyzed here were only from people who planned to

gather with other people for the 2020 anksgiving

holiday, suggesting that their concern was lower than

recommended by public health authorities at the time.

erefore, we treat an increase in concern as a positive

outcome in this study. at said, it is worth noting that

an increase in concern is not always a desirable outcome.

More generally, we hope that participants understood the

information presented and updated their beliefs based on

that information.

To compare post-intervention concern across interven

tion conditions, we ran an ANCOVA with condition as

the predictor and pre-intervention concern as a covari

ate.e Static Condition servedas the reference group for

the condition comparisons (see Fig.5a and Table2). We

found evidence that the Static Condition reported lower

concern post-intervention than the Anecdote Condition

(p < 0.001), and marginal evidence that the Static Con

dition reported lower concern post-intervention than

both the Static + Anecdote (p = 0.008) and Narrative

Visualization (p = 0.009) Conditions. e distributions of

pre-intervention concern, post-intervention concern, and

change in concerns are available in Fig.6. We would like

to note that many of the participants did not update their

concern score from pre- to post-intervention (see Explor

atory Analysis for more detail).

Next, we ran a series of one-sample t-tests to deter

mine if participants’ change in concern (posttest concern

score—pretest concern score) was signiﬁcantly diﬀerent

Table 1 Sample sizes and demographic features of samples included in the analysis

Anecdote

Condition

Static + Anecdote

Condition

Static

Condition

Narrative

Visualization

Condition

Total N Age M(SD) Gender

Pre-Thanksgiving (Session 1) 135 150 131 116 532 33.91(11.24) 242 F

Post-Thanksgiving (Session 2) 104 108 104 91 407 34.02(11.33) 201 F

Page 9 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

from 0 in each condition. All model output is available in

Table2. We found that participants in the Static Condi

tion were less concerned after viewing the intervention

materials. is decrease is reminiscent of a "backﬁre

eﬀect" (Lewandowsky et al., 2012). Past research has

shown that attempts to debunk misinformation often

lead to the denial of that information and further polar

ize opinion (Cook & Lewandowsky, 2016). When people

do not have access to information about the data-gener

ating process they may (understandably) ﬁnd it diﬃcult

to understand the data, which may ultimately lead to

reduced concern. However, it is worth noting that there

is currently debate as to whether such backﬁre eﬀects are

robust (see Swire-ompson etal., 2020 for a review).

Using persons as eﬀect sizes to achieve converging evi

dence (Grice etal., 2020), we examined the proportion

of participants per condition whose concern increased

after viewing the intervention materials (Fig. 5b). e

Static Condition was least eﬀective at increasing peo

ple’s COVID-19 concern. e proportion of people who

increased their concern was similar across the other con

ditions, with the Anecdote Condition having the largest

proportion of participants whose concern increased.

Change in Perceived Risk. It is important to note that all

participants were given explicit instructions on the risk

Fig. 5 Study 1 Change in Concern and Perceived Risk by Condition. Note. Figure 5a illustrates mean and standard error change in concern

by condition (

postconcern

−

preconcern

). A change in of zero indicates no change in concern; positive scores indicate increased concern,

and negative scores indicate decreased concern. Figure 5c illustrates mean and standard error change in perceived risk (

postrisk

−

prerisk

Figures 5a and 5c are not on the same scale. Figures 5b and 5d illustrate the proportion of participants whose concern or perceived risk increased

after viewing their respective intervention

Page 10 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

Table 2 Change in Concern and Perceived Risk from Pre- to Post-Intervention in Session 1

*indicates signicance of p < 0.003

ANCOVA Results—Session 1 (Post-Concern ~ Condition + Pre-Concern)

Coecients Estimate t value p

Intercept 0.73 0.41 0.68

Static | Static + Anecdote 5.26 2.67 0.008

Static | Narrative Visualization 5.50 2.61 0.009

Static | Anecdote 8.28 4.09 < 0.001*

Pre-intervention concern 0.87 36.54 < 0.001*

One-sample t tests—Change in concern versus 0

Condition df t p 95%CI

Static 130 − 3.16 0.002* [− 7.74, -1.78]

Anecdote 134 2.46 0.02 [0.79, 7.21]

Static + Anecdote 149 0.52 0.60 [− 1.75, 2.99]

Narrative Visualization 115 0.73 0.47 [− 1.96, 4.27]

ANCOVA Results—Session 1 (Post-Risk ~ Condition + Pre-Risk)

Coecients Estimate t value p

Intercept 25.90 14.18 < 0.001

Static | Static + Anecdote − 2.80 − 1.29 0.20

Static | Narrative Visualization 3.01 1.30 0.19

Static | Anecdote 3.36 1.51 0.13

Pre-Intervention Risk 0.45 15.79 < 0.001*

One-sample t tests—Change in Risk versus 0

Condition df t p 95%CI

Static 130 3.30 0.001* [3.23, 12.93]

Anecdote 134 8.57 < 0.001* [10.25,

16.40]

Static + Anecdote 149 3.30 0.001* [2.58, 10.27]

Narrative Visualization 115 6.19 < 0.001* [9.53, 18.49]

Fig. 6 Distributions of Concern Scores

Page 11 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

of dining with 10 people, regardless of condition. us,

the risk question asked post-intervention -“What is the

risk that at least one person at a anksgiving table with

10 people has COVID-19?” - was a simple recall task.

We compared post-intervention perceived risk between

conditionsusing anANCOVA, controlling for pre-inter

vention risk perception as a covariate. Pre-intervention

perceived risk predicted post-intervention perceived risk

(p < 0.001); however, there was no eﬀect of condition on

perceived risk (see Table2).

One-sample t-tests with the test value as 0 (no change

in perceived risk) indicated that perceived risk increased

among all conditions after the intervention. Using per

sons as eﬀect sizes to achieve converging evidence (Grice

etal., 2020), we examined the proportion of participants

per condition whose perceived risk increased after view

ing the intervention materials (Fig.5d). e Static Condi-

tion was theleast eﬀective at increasing perceived risk,

although proportions were similar across conditions (see

Table2).

Concern about anksgiving Plans. After complet

ing the intervention, participants were also asked to rate

how concerned they were with their current anksgiv

ing plans on a scale of 0–100 (not concerned—extremely

concerned). A one-way ANOVA did not ﬁnd a diﬀerence

in concernbetween the four conditions (Mean Ratings:

Static Condition = 40.09, Narrative Visualization Condi

tion = 39.81, Static + Anecdote Condition = 42.66, Anec-

dote Condition = 43.91; see Table3).

Post‑thanksgiving follow‑up survey

We next examined the data from the participants who

returned for Session 2 and completed the Post-anks

giving survey. We examined data from participants who

indicated that they planned to spend anksgiving with

others during the Pre-anksgiving session (n = 407)

as we considered those participants to be the least

risk-averse.

Holiday Plans. Participants were asked to report infor

mation on how they spent their anksgiving holiday. We

compared the proportion of participants who reported

making last-minute changes to their anksgiving plans

(n = 80) between conditions with a Chi-square test and

found no diﬀerences between conditions (see Table3).

Participants who did change their anksgiving plans

were asked whether the change increased, decreased, or

did not impact the number of people at their gathering.

e proportion of participants who  spent the holiday

with fewer people was compared between conditions,

and there was no signiﬁcant diﬀerence between condi

tions (see Table3).

Participants were also asked whether they planned to

spend the December/January holidays alone or with oth

ers. We excluded participants who responded “Unsure”

(n = 22) and compared the proportion of participants

who planned to spend the holidays with others between

the four conditions usinga Chi-square test. ere was no

signiﬁcant diﬀerence between groups, with 54% of the

Static + Anecdote Condition, 59% of the Narrative Visu

alization Condition, 62% of the Static Condition, and 51%

of the Anecdote Condition, indicating that they planned

to spend the holidays with others (see Table3).

Participants were also asked whether the COVID pan

demic had changed their usual holiday plans. We com-

pared the proportion of participants from each condition

who indicated that they planned to spend the holidays

with fewer people than usual using a Chi-square test.

ere was no signiﬁcant diﬀerence between conditions,

with 69% of the Static + Anecdote Condition, 80% of the

Narrative Visualization Condition, 69% of the Static Con

dition, and 80% of the Anecdote Condition indicating

that they planned to spend the holidays with fewer peo

ple than usual (see Table3).

Table 3 Behavioral Intentions Reported in Sessions 1 and 2

Pre-Thanksgiving Intentions

A. ANOVA—Comfort with Current Thanksgiving Plans

F (3,528) p

Overall Model 0.7 0.55

Post-Thanksgiving Intentions

2 p

Proportion of participants who changed Thanksgiving plans by condition 3.13 0.37

Proportion of participants who decreased Thanksgiving attendees by condition 1.37 0.71

Proportion of participants who planned to spend holidays with others by condition 3.28 0.35

Proportion of participants who planned to spend holidays with fewer people by condition 6.07 0.10

Page 12 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

Concern Ratings. We also asked participants how con-

cerned they were about getting COVID or their family

getting COVID at holiday gatherings. eir responses to

these two items were averaged to create a composite con

cern score. We compared average concern between con-

ditions with a one-way ANOVA and found no signiﬁcant

diﬀerences in concern between conditions (see Table4).

We did not use an ANCOVA as in Session 1 since we

did not have participants rate their concern about the

December holidays pre-intervention.

Change in Perceived Risk. Participants asked to report

the risk (0–100%) that at least one person at a table of 10

would have COVID at a anksgiving dinner. We investi

gated change in perceived risk using an ANCOVA model

with condition as the predictor and pre-intervention

perceived risk as the covariate. We found no signiﬁcant

diﬀerences between groups (see Table 4). One-sample

t-tests compared thechange in perceived risk from pre-

intervention to Session 2 to 0 for each condition, and

we did not ﬁnd evidence for signiﬁcant increases in per

ceived risk for any of the conditions (see Table4).

Exploratory analyses

We ran a series of exploratory analyses to investigate (1)

the factors leading to participants’ decisions and concern

relating to real-world plans, and (2) the factors leading

participants to change their concern/perceived risk after

viewing the intervention materials. Since these analyses

are exploratory, we decided to be more lenient with our

interpretation of signiﬁcance, setting our alpha level to

p < 0.05 (Fig.7).

Exploratory Mediation Models. As reported above, we

did not ﬁnd an eﬀect of condition on participants’ real-

world behavioral intentions. us, we explored the pos

sibility that post-intervention concern mediated the

relationship between condition and real-world behav

ioral intentions and decision-making. We ran a series

of mediation analyses with condition as the predictor,

pre-intervention concern as a covariate, post-interven

tion concern as the mediator, and real-world decisions/

concern about plans as the outcome variables. For each

model, only two conditions were included, with the

Static Condition as the reference group, considering that

the Static Condition showed a backﬁre eﬀect in Session

1. Using the Static Condition as the reference condi

tion also allowed us to investigate the impact of adding

an anecdote (Static + Anecdote) or explanation of the

data (Narrative Visualization) to data on concern about

plans/behavioral intentions. e speciﬁc outcome vari

ables used were: (1) concern about current anksgiving

plans (collected in Session 1), (2) whether participants

actually changed their anksgiving plans after the inter

vention (collected in Session 2), and (3) whether par-

ticipants planned to spend the December holidays with

others (collected in Session 2). All of the mediation mod

els were run using the PROCESS macro in R (Hayes,

2017), with the Model 4 speciﬁcation. We did not ﬁnd

any evidence that post-intervention concern mediated

the relationship between condition and whether par

ticipants actually changed their anksgiving plans, nor

whether participants planned to spend the December

holidays with others. However, we found evidence for a

Table 4 Concern and Change in Perceived Risk in Session 2

*indicates signicance of p < .003

F (3403) p

Overall Model 1.08 0.36

ANCOVA Results—Session 1 (Post-Risk ~ Condition + Pre-Risk)

Coecients Estimate t value p

Intercept 28.76 10.74 < 0.001

Static | Static + Anecdote 1.43 0.44 0.66

Static | Narrative Visualization − 0.31 − 0.09 0.93

Static | Anecdote 0.65 0.20 0.84

Pre-Intervention Risk 0.17 3.88 < 0.001*

One-sample t tests—Change in Risk versus 0

Condition df t p 95%CI

Static 103 1.29 0.20 [− 2.18, 10.30]

Anecdote 103 1.49 0.14 [− 1.61, 11.36]

Static + Anecdote 107 1.74 0.08 [− 0.77, 12.03]

Narrative Visualization 90 1.99 0.05 [0.02, 13.54]

Page 13 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

full mediation of post-intervention concern on the rela-

tionship between condition and participants’ concern

with their current anksgiving plans in Session 1. is

evidence was present for all three models comparing the

Static Condition to the rest of the intervention conditions

(see Fig. 8). ese ﬁndings suggest that relative to the

Static Condition, the other interventions led to increased

concern, which in turn increased participants’ concerns

about their anksgiving plans. Given that all partici

pants planned to celebrate anksgiving with others,

this suggests that participating in the three interventions

impacted their risk perception of real-world decisions in

a positive way relative to the Static Condition.

Predicting Increased Concern. As illustrated by Fig.6,

it was surprising to us that many participants’ concerns

did not increase after viewing their respective interven

tion. us, we ran an exploratory analysis examining the

factors predicting whether one’s concern increased after

viewing an intervention. We ran a logistic regression

with political partisanship (coded as right, left, or other,

with right as the reference group), numeracy based on

the subjective numeracy scale, condition (with the Static

Condition as the reference group), and pre-intervention

Fig. 7 Distributions of Perceived Risk

Fig. 8 Exploratory Mediation Models. Note: For each model, the Static Condition is used as the reference group for each Condition comparison.

*p < .05, *p < .001

Page 14 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

concern as predictors, and whether participants’ concern

increased (1) or stayed the same/decreased (0) as the out

come variable:

We found evidence that participants in all three condi

tions were more likely to increase their level of concern

than the Static Condition, and that participants who

identiﬁed with the political left (i.e., Democrat or Lean

Democrat) were more likely to increase their level of con

cern than participants who identiﬁed with the political

right (i.e., Republican or Lean Republican) (see Table5).

ere was no eﬀect of numeracy nor pre-intervention

concern on whether participants’ concern increased.

Discussion

e goal of Study 1 was to examine whether narrative

visualizations are an eﬀective tool for communicating

cumulative risk, as well ashow other interventions such

as the use of anecdotes and static data visualizations

inﬂuenced perceptions of cumulative risk. We collected

these data in a unique scenario where participants rea

soned about the real-world risks associated with dining

with others during the COVID pandemic in 2020. Our

most notable ﬁnding was a data backﬁre eﬀect in the

Static Condition in Session 1, suggesting that viewing the

cumulative risk data led to decreased concern about the

risk of contracting COVID at anksgiving dinner. It is

important to note that this lack of concern was not due

toa misunderstandingof the risk information presented

to participants. Indeed, participants across all conditions

were presented with the same cumulative risk of con

tracting COVID from a 10-person dinner, and there were

no signiﬁcant diﬀerences in perceived risk post-interven

tion between conditions. is suggests that although par-

ticipants in the Static Condition perceived risk similarly

Increased Concern

(

)

∼

Condition

Partisanship

Numeracy

Pre

−

Concern

to the other conditions, they were simply less concerned

about the risk.

Table 5 Factors associated with concern increase—logistic

regression output

Estimate Std. Error z value p

Intercept − 1.10 0.49 − 2.24 0.03

Static + Anecdote versus Static 0.60 0.25 2.40 0.02*

Narrative Visualization ver-

sus Static

0.64 0.27 2.38 0.02*

Anecdote versus Static 0.80 0.26 3.12 0.002*

Partisanship: Left versus Right 0.53 0.23 2.30 0.02*

Partisanship: Other versus Right 0.07 0.26 0.24 0.80

Numeracy 0.03 0.08 0.39 0.70

Pre-intervention concern − 0.003 0.003 − 0.99 0.32

Four weeks after the intervention, we did not ﬁnd statis-

tically signiﬁcant diﬀerences between conditions in terms

of their behavioral intentions. A similar proportion of par

ticipants reported changing their anksgiving plans after

the intervention among conditions, and the proportion

of participants who planned to spend the holidays alone

or with fewer people was also similar across conditions.

We found little evidence that the eﬀects of the interven

tion were long lasting, as concern ratings and change in

perceived risk were similar across conditions at the Post-

anksgiving survey stage. However, in an exploratory

mediation analysis we foundevidence that relative to the

Static Condition, participating in the Static + Anecdote,

Narrative Visualization, and Anecdote Condition was

associated withgreatergeneral concernwhich was asso

ciated with increasedconcern about one’s anksgiving

plans to dine with others. is provides some evidence

that participants connected the data presented to them to

their real-world decisions and intentions.

Lastly, we found that a large number of participants

did not report any increase in concern from pre- to post-

intervention. In an exploratory analysis, we found evi

dence that participants who politically leaned to the left

were more likely to increase their concern from pre- to

post-intervention compared to those who leaned right.

is suggests that left-leaning participants may have

been more inclined to update their beliefs in the presence

of new information.

e most notable ﬁnding from Study 1 was that view

ing data by itself led to a backﬁre eﬀect—where partici-

pants were less concerned about a risk after viewing a

bar graph illustrating cumulative risk. However, viewing

static data in conjunction with an anecdote prevented

this data backﬁre eﬀect, as did viewing a narrative vis

ualization. In Study 2, we further explore why these

interventions did not lead to a backﬁre eﬀect by exam

ining how the various interventions inﬂuenced trust and

understanding of the data.

Study 2

Study 1 showed that viewing a static visualization of

cumulative risk was associated with increased risk per

ception but decreased concern from pre- to post-inter-

vention. is backﬁre eﬀect was not present inthe other

conditions where participantsviewed either an anecdote

by itself, a static visualization with an anecdote, or a

Page 15 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

narrative visualization. is provides promising evidence

that interventions can include data without leading to

backﬁre eﬀects on concern. It is also encouraging that the

Narrative Visualization Condition did not lead to a data

backﬁre eﬀect even though this intervention did not rely

on emotionalappeal like the other conditions (Anecdote/

Static + Anecdote Conditions), suggesting that narrative

visualizations may be an eﬀective tool for communicating

risks. Why didn’t the Narrative Visualization Condition

lead to the data backﬁre eﬀect that occurred in the Static

Condition? It is possible that people who viewed the nar

rative visualization were able to obtain a better under-

standing and/or had more trust in the data because of

the additional information provided by the walkthrough,

compared to the static visualization. e aim of Study 2

was to replicate our ﬁndings from Study 1 while examin

ing whether the additional information provided by the

narrative visualization aﬀected trust, subjective under

standing, or both. We also examined how these variables

were inﬂuenced by the presence of anecdotes, which do

not include information about how the data were gener

ated, but also mitigated thebackﬁre eﬀectin Study 1. All

data, materials, and code are available at https:// osf. io/

k43ev/? view_ only= 67cc6 401b0 69468 89ac4 99fc1 9afec 2f.

Methods

Participants

We recruited 1592 US participants (806 female; age

M(SD) = 44.67(16.04)) from Proliﬁc between January 6,

2022 and February 5, 2022. Participants received an aver

age pay of $17.04/hr (USD), and the median completion

time for the study was 5min and 36s. All procedures

were determined to be exempt by the University of Mich

igan IRB.

Design

As in Study 1, participants were randomly assigned to

one of four conditions in which they viewed either a static

visualization (Static Condition, n = 399), the same static

visualization combined with an anecdote (Static + Anec

dote Condition, n = 388), an anecdote presented without

a visualization (Anecdote Condition, n = 404), or a nar

rative visualization (Narrative Visualization Condition,

n = 401).

Materials

Pre-Intervention Materials. Data for Study 2 were col-

lected in January and February of 2022, so we made

severalchanges to the Study 1 materials to reﬂect devel

opments that had occurred in the world (e.g., the devel-

opment of COVID-19 vaccines) during the time between

the two studies. To account for changes in the COVID-19

pandemic and access to vaccines, participants were asked

to make judgments about their hypothetical anksgiv

ing plans during a pandemic with a ﬁctional disease for

which there was not yet a vaccine (as there were no vac

cines available for COVID-19 when the Study 1 data were

collected). All items with exact wording are included in

the Supplementary Materials. All participants were pre

sented with this scenario:

Imagine the following scenario: a new respiratory

disease emerged in your community. is new dis

ease is highly contagious, and there is no vaccine

available yet. e Fall & Winter holiday season is

approaching, and public health oﬃcials recommend

that people do not gather during the holidays, but

your local government has not placed any restric

tions on gathering.

Please answer some questions about what your

anksgiving plans would have been this year if you

were living through the above hypothetical sce

nario.

Next, participants were asked who they would have

anksgiving dinner with in this hypothetical scenario

(i.e., alone/with people who live with you, extended fam

ily, friends/neighbors, strangers). ey were also asked to

indicate how concerned they would be about contracting

the disease at anksgiving dinner and how concerned

they would be that someone else at their anksgiving

dinner would contract the disease (slider scales from 0

(not concerned at all)—100 (extremely concerned) for

each item). To assess understanding of risk, participants

were asked “If 4% of the population has the disease, what

is the risk that at least one person at a anksgiving table

with 10 people has the disease?” which they answered

with a slider scale from 0 to 100%.

Intervention Materials. We made several reﬁnements

to the Study 1 intervention materials for Study 2. In the

Static and Static + Anecdote Conditions of Study 2, we

presented the risk information as an icon array (Fig.9),

rather than the bar graph from Study 1 (Fig. 3). Icon

arrays are thought to increase understanding of risk, par

ticularly for individuals with low numeracy (Okan etal.,

2015). We made this change to improve the robustness of

the Static Condition as a baseline. In addition, the narra

tive visualization in Study 1 used seven consecutive dis-

plays (Fig.1), while the static intervention used a single

static visualization (Fig.3). To reduce the diﬀerence in

time on task between these two interventions, we short

ened the narrative visualization to just three displays

(Fig.10). Finally, the anecdote intervention in Study 1 did

not include the quantitative risk data presented in the

other interventions. To reduce diﬀerences between inter

ventions, we modiﬁed the ﬁnal paragraph of the anecdote

Page 16 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

to include the critical quantitative conclusion below

(bold text was added for Study 2):

I’m so scared, and words can’t describe the guilt I

feel. I know I’m just one person, but if you take any

thing away from this story please STAY HOME. e

pandemic has been hard for everyone, and we

all miss our families, but the risks can be much

higher than we realize. I found out afterward

that (according to the CDC) even though only

about 4 out of 100 people are infected, there was

about a 40% chance that someone at our dinner

of 10 was infected. So please don’t risk something

like this happening to you or your family. Stay home

not just to protect yourself, but to protect your loved

Fig. 9 Static + Anecdote Condition materials for Study 2. Note. The Static Condition materials contained the same icon array without the anecdote.

The Anecdote Condition materials contained an expanded version of the anecdote with no visualized data (see Supplementary Materials)

Fig. 10 Study 2 Narrative Visualization Materials. Note. Participants viewed three consecutive pages which explain step-by-step how the risk

of exposure to the disease is about 40% in the study scenario (i.e., a Thanksgiving dinner with 10 people taking place in a community where 4

out of 100 people are infected)

Page 17 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

ones too.

We also modiﬁed the anecdote to remove any lan

guage that referred to COVID speciﬁcally. e full anec-

dote used in Study 2 is available in the Supplementary

Materials.

Post-Intervention Survey Materials. After viewing the

intervention materials, we asked participants to rate

their concern about contracting the disease at anksgiv

ing dinner for themselves and for others on slider scales

of 0 (not concerned at all) to 100 (extremely concerned).

We also asked them how they felt about the hypotheti

cal anksgiving plans they reported pre-intervention

with a slider scale from 0 (not concerned at all) to 100

(extremely concerned).

To measure self-rated understanding, we asked partici

pants to rate their level of agreement with the following

statement from 0 to 100: “After viewing the data, I under

stand why a anksgiving dinner with 10 people has

about a 40% chance of exposure to the disease.” To meas

ure trust, we asked them to rate their level of agreement

on scales of 0–100 with the following items:

1. I feel like the data are intended to accurately portray

the risks of the new disease.

2. I feel like the data do accurately portray the risks of

the new disease.

Lastly, participants provided demographic information,

including numeracy and political partisanship data as in

Study 1.

Procedure

e procedure for Study 2 was the same as for Study 1

except thatparticipants were not invited for a follow-up

survey four weeks later.

Results

As in Study 1, given that we present the results of mul-

tiple tests, we used a Bonferroni correction to reduce

the likelihood of Type I error. erefore, in the following

analyses comparing concern, trust, and understanding

between conditions, we lowered our signiﬁcance level to

p < 0.007, as 7 of our tests did not correct for multiple

comparisons (

.05

.007

Exclusion criteria

e Post-Intervention Survey contained two attention

check items:

1. According to the data at the end of the story, what is

the approximate risk that you will be exposed to the

disease if 4 out of 100 people in your community are

infected and you have dinner with 10 people? (40%,

4%, 10%, or 14%?)

2. In the hypothetical scenario you read at the begin

ning of the study, was a vaccine available for the dis-

ease? (Yes, one vaccine, Yes, multiple vaccines, or

No).

Participants were excluded from the analysis if they

gave an incorrect response on one or both attention

check questions (n = 156). Of participants who passed

both attention check items, only 12% of participants

(n = 173) stated that they would eat dinner with people

outside of their home, so we were unable to divide the

data based on dinner intentions as we did for Study 1.

Post‑intervention concern

All model output is available in Table6. To compare post-

intervention concern across intervention conditions, we

ran an ANCOVA with condition as the predictor, pre-

intervention concern as the covariate, and post-inter

vention concern as the outcome variable of interest (see

Fig.11a). Unlike Study 1, there were no signiﬁcant diﬀer

ences between conditions.

Next, we ran a series of one-sample t-tests with the

comparison value as 0 to determine if change in con

cern signiﬁcantly increased or decreased in each condi-

tion. After correcting for multiple tests, we did not ﬁnd

evidence for a signiﬁcant increase or decrease in partici

pants’ concern in any of the conditions. Using persons as

eﬀect sizes to achieve converging evidence (Grice etal.,

2020), we examined the proportion of participants per

condition whose concern increased after viewing the

intervention materials (Fig.11b). Similar to Study 1, the

Static Condition was least eﬀective at increasing people’s

COVID-19 concern. Distributions of pre-intervention,

post-intervention, and change in concern scores are

shown in Fig.12.

The eect ofcondition ontrust andunderstanding

Our primary question in Study 2 was whether the nar-

rative visualization intervention and Static + Anecdote

Conditions prevented a data backﬁre eﬀect because par

ticipants felt that they understood and trusted the data

more than in the Static Condition. We ﬁrst tested our

hypothesis that viewing the narrative visualization would

improve understanding and trust compared to the Static

Condition.

Self-rated understanding was measured through

response to the item “After reading the story, I under

stand why a anksgiving dinner with 10 people has

Page 18 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

about a 40% chance of exposure to the disease.” To exam-

ine the role of condition on self-rated understanding, we

ran a one-way ANOVA with Condition as the predictor

and understanding as the outcome variable. e model

was signiﬁcant overall (see Fig. 13a; Table 7) and post

hoc tests with Tukey HSD indicated that participants in

the Narrative Visualization Condition reported greater

understanding than those inthe Static + Anecdote Con

dition, the Anecdote Condition, and the Static Condition.

Trust was calculated by averaging participants’

responses to the two items: (1) "I feel like the story is

intended to accurately portray the risks of the new dis

ease" and (2) "I feel like the story does accurately portray

the risks of the new disease". We again ran a one-way

ANOVA with condition as the predictor and trust as the

outcome variable. e model was signiﬁcant overall (see

Fig. 13b; Table 7) and post hoc tests with Tukey HSD

indicated that participants in the Narrative Visualization

Condition reported greater trust than those inthe Static

Condition.

Mediation analysis

e primary goal of Study 2 was to investigate the pos-

sible mechanisms that allowed participants in the Nar-

rative Visualization Condition to view the data without

experiencing a data backﬁre eﬀect (as seen in the Static

Table 6 Change in Concern from Pre- to Post-Intervention in Study 2

*Indicates signicance of p < .007

ANCOVA Results—Study 2 (Post-Concern ~ Condition + Pre-Concern)

Coecients Estimate t value p

Intercept 11.99 7.68 < 0.001

Static | Static + Anecdote 3.34 2.29 0.02

Static | Narrative Visualization 2.35 1.63 0.10

Static | Anecdote 0.92 0.63 0.53

Pre-intervention concern 0.80 45.53 < 0.001*

One-sample t tests—Change in concern versus 0

Condition df t p 95%CI

Static 356 − 0.95 0.34 [− 3.06, 1.06]

Anecdote 354 − 0.44 0.66 [− 3.01, 1.91]

Static + Anecdote 350 2.26 0.02 [0.30, 4.31]

Narrative Visualization 372 1.20 0.23 [− 0.73, 2.99]

Fig. 11 Study 2 Changes in Concern and Perceived Risk by Condition. Note. Figure 11a illustrates condition means and standard error. A change

in concern of zero indicates no change; positive scores indicate increased concern, and negative scores indicate decreased concern. Figure 11b

shows the proportion of participants whose concern increased in each condition

Page 19 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

Condition). As such, we investigated the impact of inter-

vention condition on participants’ self-reported under-

standing of the data and trust in the data.

Mediation analysis—comparing narrative visualization

andstatic conditions

We conducted a mediation analysis using the PROCESS

macro in R (Hayes, 2017) to determine whether the

impact of the narrative visualization on post-concern rat

ings was serially mediated by understanding and trust

in the data (Model 6 in the PROCESS macro). We ran

a mediation analysis with understanding and trust as

mediators on the eﬀect of the narrative visualization (IV)

on post-intervention concern (DV) (Fig.14). To isolate

the speciﬁc eﬀect of the narrative visualization, we com

pared the narrative visualization intervention to the static

intervention. Pre-intervention concern was included as a

covariate.

e Eﬀect of Condition on Understanding and Trust.

We ran a serial model with the idea that understanding

aﬀects trust. e outcome is shown in Fig.14. We found

evidence for an independent eﬀect of condition on under

standing: participants in the Narrative Visualization

Condition reported greater understanding compared to

Fig. 12 Distributions of concern scores

Fig. 13 Study 2 self-rated understanding and trust by condition. Note. Figures 13a–b illustrate mean and standard error for self-rated understanding

and trust by condition

Page 20 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

participants in the Static Condition (b = 12.12, p < 0.001,

95% CI = [8.00, 16.24]). ere was also an independ

ent eﬀect of understanding on trust, such that greater

understanding was associated with higher trust (b = 0.75,

p < 0.001, 95% CI = [0.70, 0.79]). However, there was no

signiﬁcant independent eﬀect of condition on trust itself

(b = − 0.65, p = 0.64, 95% CI = [− 3.41, 2.10]).

Predicting Post-Intervention Concern. ere was an

independent eﬀect of trust in the data on concern about

disease transmission post-intervention (b = 0.09, p = 0.01,

Table 7 Trust and understanding by condition in Study 2

*Indicates signicance of p < .007

Study 2—Understanding

ANOVA—Understanding by Condition

F (3,1432) p

Overall Model 21.16 < 0.001*

Tukey HSD Signicant Pairwise Comparisons

di

adj

95%CI

Narrative Visualization | Static 13.48 < 0.001* [7.37, 19.59]

Anecdote | Static + Anecdote − 4.41 0.26 [− 10.59, 1.77]

Anecdote | Narrative Visualization − 17.89 < 0.001* [− 23.98, − 11.80]

Static | Static + Anecdote 1.08 0.97 [− 5.09, 7.26]

Static | Narrative Visualization − 12.39 < 0.001* [− 18.48, − 6.31]

Anecdote | Static − 5.49 0.10 [− 11.65, 0.67]

Study 2—Trust

ANOVA—Trust by Condition

F (3,1432) p

Overall Model 6.23 < 0.001*

Tukey HSD Signicant Pairwise Comparisons

di

adj

95%CI

Narrative Visualization | Static 6.28 0.02 [0.59, 11.97]

Anecdote | Static + Anecdote 3.68 0.36 [− 2.08, 9.43]

Anecdote | Narrative Visualization − 2.60 0.64 [− 8.28, 3.07]

Static | Static + Anecdote − 2.53 0.67 [− 8.28, 3.22]

Static | Narrative Visualization − 8.81 < 0.001* [− 14.47, − 3.14]

Anecdote | Static 6.20 0.03 [0.47, 11.94]

Fig. 14 Study 2: Mediation model comparing the Narrative Visualization Condition to the Static Condition. Notes: *p < 0.05; **p < 0.001. Thick lines

show the signiﬁcant indirect pathway from condition to post-intervention concern via understanding and trust

Page 21 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

95%CI = [0.02, 0.17]). However, reported understanding

did not independently predict post-intervention concern

(b = − 0.005, p = 0.90, 95% CI = [− 0.08, 0.07]). ere was

also no evidence of a direct eﬀect of condition on post-

intervention concern (b = 1.57 p = 0.25, 95% CI = [− 1.15,

4.30]).

Mediating the Relationship Between Condition and

Concern. We did not ﬁnd evidence for an indirect eﬀect

of condition on concern through understanding (indi

rect eﬀect = − 0.05, SE = 0.44, 95% CI = [− 0.94, 0.82]) or

trust (indirect eﬀect = − 0.06, SE = 0.15, 95% CI = [− 0.38,

0.21]). However, the data suggested that condition had

an indirect eﬀect on concern through both self-rated

understanding and trust (indirect eﬀect = 0.85, SE = 0.37,

95% CI = [0.20, 1.64]). ese ﬁndings suggest a full serial

mediation of understanding and trust on the relationship

between condition and post-intervention concern.

Mediation analysis—comparing static + anecdote andstatic

conditions

In Study 1, we found evidence of a data backﬁre eﬀect,

in which participants in the Static Condition reported

feeling less concerned after seeing the data. Even though

participants in the Static + Anecdote and Narrative Visu

alization Conditions alsoviewed data, they did not expe-

rience the same eﬀect. erefore, we thought it would be

helpful to examine the mechanisms that prevented the

data backﬁre eﬀect in the Static + Anecdote Condition.

We again conducted a mediation analysis toexamine the

eﬀect of condition, understanding, and trust on post-

intervention concern.

e Eﬀect of Condition on Understanding and Trust.

We did not ﬁnd an independent eﬀect of condi

tion on reported understanding (b = −1.14, p = 0.65,

95% CI = [− 5.95, 3.67]); however, participants in the

Static + Anecdote Condition reported feeling higher

levels of trust than participants in the Static Condi

tion (b = 3.28, p = 0.02, 95% CI = [0.58, 5.97]). As in

the previous analysis, there was an independent eﬀect

of understanding on trust; participants who reported

greater understanding also reported greater trust in the

data (b = 0.72, p < 0.001, 95% CI = [0.12, 0.22]).

Predicting Post-Intervention Concern. We found evi

dence for a direct eﬀect of condition on post-intervention

concern, such that participants in the Static + Anecdote

Condition reported greater concern than the Static Con

dition (b = 3.15, p = 0.03, 95% CI = [0.40, 5.91]). ere

was also an independent eﬀect of trust on concern,

with participants reporting greater trust also reporting

greater concern post-intervention (b = 0.08 p = 0.03, 95%

CI = [0.007, 0.16]). ere was no independent eﬀect of

reported understanding on post-intervention concern

(b = 0.02, p = 0.63, 95% CI = [− 0.05, 0.08]).

Mediating the Relationship Between Condition and

Concern. We did not ﬁnd evidence for an indirect eﬀect

of condition on concern through understanding (indirect

eﬀect = − 0.02, SE = 0.11, 95% CI = [− 0.27, 0.21]) or trust

(indirect eﬀect = 0.27, SE = 0.20, 95% CI = [− 0.005, 0.74]).

Unlike the previous analysis, we found no evidence of

an indirect eﬀect of condition on post-intervention con

cern through both understanding and trust (indirect

eﬀect = − 0.07, SE = 0.17, 95% CI = [− 0.45, 0.25] (see

Fig.15).

Discussion

Unlike inStudy 1, we did not observe statistically sig-

niﬁcant diﬀerences between intervention conditions in

changes in concern. Study 2 results likely diﬀered from

Study 1 because of the diﬀerences in context between

the two studies: Study 1 assessed concern about real-

life plans during a time where COVID-19 was much

less understood, while Study 2 asked about concern for

hypothetical plans and was administered after the virus

was better understood and a vaccine was widely available.

Additionally, Study 2 includeddata from all the partici

pants, whereas Study 1 exclusively used data from par-

ticipants who were the least risk averse (indicated they

Fig. 15 Study 2: Mediation model comparing the Static + Anecdote Condition to the Static Condition. Notes: *p < 0.05; **p < 0.001

Page 22 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

planned to spend anksgiving with people outside of

their household).

e ﬁndings from Study 2 provide a mechanistic expla

nation for why participants in the Narrative Visualization

Condition did not experience a backﬁre eﬀect observed

in the Static Condition. Our data suggest that viewing a

narrative visualization led to increased subjective under

standing of data and greater trust. Most importantly,

a mediation analysis indicated that viewing the narra

tive visualization increased concern relative to viewing

a static visualization, and that this eﬀect was produced

through increasing understanding and trust in the data.

us, the model shows that the narrative visualization

helped people feel more conﬁdent in their understand

ing of the risk, which, in turn, increased their trust in the

information,ultimately resulting in an increase in overall

concern.

General discussion

Helping the general public accurately reason about data,

especially in contexts related to real-world decision-mak

ing, is an important goal of science communication. e

current investigation builds on previous work examining

how viewing quantitative data and reading aﬀect-laden

stories shape risk perception across various domains

(Cutello etal., 2020; Fagerlin etal., 2005; Rakow etal.,

2015; Tiede etal., 2022). In Study 1, we compared diﬀer

ent methods for presenting cumulative risk data regard-

ing the likelihood of contracting COVID during the 2020

holidays. We tested the eﬀectiveness of narrative visuali

zations, a method of data communication that walks par-

ticipants through how data are generated, as a scientiﬁc

communication tool. We also examined the impact of

static data visualizations and anecdotal evidence on risk

perception and concern about the COVID pandemic.

e context ofstudy allowed us to investigate how these

interventions aﬀected individuals with suboptimal risk

perception or concern, in that participants included in

our data analysis planned to dine with people outside

of their household during the 2020 American anks

giving holiday. We were also able to examine the eﬀects

of viewing the intervention materials on perceptions of

real-world decisions given the longitudinal design ofin

Study 1. Akey ﬁnding from Study 1 was that participants

who viewed a static visualization of cumulative risk expe

rienced a backﬁre eﬀect—even though their perceived

risk of dining with individuals outside of their home

increased from pre- to post-intervention, their concern

about contracting COVID from pre- to post-intervention

actually decreased. is backﬁre eﬀect was not present

observed in the Static + Anecdote, or the Narrative Visu

alization Conditions.

It is perhaps unsurprising that the presence of an anec

dote in the Static + Anecdote Condition helped mitigate

this backﬁre eﬀect, given prior work demonstrating the

potential for emotion-laden stories to increase perceived

risk (Freling etal., 2020). Frequently, risk avoidance strat

egies rely on providing people with alarming stories. For

instance, to avoid driving distracted or driving under the

inﬂuence of drugs or alcohol, risk messaging relies on

impressionable and emotional testimonials rather than

quantitative data (Betsch etal., 2011; Janssen etal., 2013;

Kim etal., 2017). Although anecdotes are often eﬀective,

they can be problematic. Typically, anecdotes rely on

aﬀect-based mechanisms activating heuristic thinking,

rather than analytic mechanisms (Rodriguez etal., 2016).

Consequently, people tend to ignore quantitative data and

instead make decisions based on testimonials or anec

dotes even when data provided are more representative.

Narrative visualizations, in contrast, do not rely on

aﬀective responses anddirectly address the data at hand.

In Study 1, we found that viewing a narrative visualiza

tion explaining how the accumulated risk data were

generated prevented a backﬁre eﬀect. is suggests that

in real-world contexts where perceived risk is subopti

mal, narrative visualizations can inﬂuence people’s atti-

tudes by communicating quantitative data, rather than

relying on stories or testimonials. Why was the nar

rative visualization eﬀective? Our mediation analysis

in Study 2 indicated that the mechanisms driving this

eﬀect wereunderstanding and trust. A serial mediation

analysis revealed a full serial mediation of understand

ing and trust on the relationship between condition

and post-intervention concern. It is perhaps unsurpris

ing that the narrative visualization increased perceived

understanding and trust; the narrative visualization

explainsthe data-generating processin detail, providing

a causal explanation for the data. Research suggests that

providing coherent explanatory frameworks is associ

ated with higher judgements of information credibility

(Sloman, 1994; agard, 2007). Causal explanations are

also more eﬀective in correcting misinformation com

pared to providing people with corrective information

(Lammers etal., 2020; Lewandowsky etal., 2012; Nyhan

& Reiﬂer, 2015; Seifert, 2002). Understanding and trust

did not mediate the relationship between condition and

concern when comparingthe Static + Anecdote and Static

Conditions, suggesting that theStatic + Anecdote Condi

tion mitigated backﬁre through a mechanism other than

increased trust and understanding.

Implications forscientic communication

Our ﬁndings provide evidence that risks can be com-

municated as quantitative data in a manner that still

increases people’s understanding and trust in data.

Page 23 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

ese ﬁndings suggest that narrative visualizations may

be an eﬀective tool for combating misinformation about

data. Over- and underestimation of risks are common

types of misinformation. Underestimation of risk, as

observed in the present studies, is particularly com

mon, and it is diﬃcult to convince individuals to follow

public health guidelines when risks are underestimated.

Overestimation of risks is also problematic. For exam

ple,vaccine-hesitant individualsoverestimate the risk of

vaccines and believethese risks are more common and

serious than they are in reality. Most prior work on cor

recting misinformation, however, focuses on correct-

ing inaccurate facts (e.g., the belief that vaccines cause

autism) rather than correcting data. For contexts such

as vaccine hesitancy, there is some evidence that correc

tions can backﬁre (Nyhan & Reiﬂer, 2015). In Study 1, we

demonstrate that correcting quantitative-and not just

factual-misinformation can also lead to backﬁre when

data are surprising. Furthermore, we show that causal

explanations are eﬀective not only for correcting misin

formation about facts, but also about data.

Our results demonstrate that narrative visualiza

tions can be more eﬀective than static visualizations for

communicating risk, particularly in situations like the

COVID-19 pandemic. Beyond this scenario, we believe

narrative visualizations have the potential to communi

cate about complex, unintuitive, or surprising data in var-

ious other contexts. In particular, narrative visualizations

may be useful in contexts involving accumulation(e.g.,

exposure to environmental toxins, or repeated engage

ment in risky behaviors) or exponential growth (com-

pound interest in the context of ﬁnancial literacy, or

the spread of wildﬁres in the context of public safety),

and in explaining outcomes related to algorithmic deci

sion-making (e.g., recidivism prediction; O’Neil, 2017).

By interleaving visualizations with explanations of the

underlying data-generating processes, narrative visuali

zations can help readers build connections between the

story and rationale,consequently increasing their trust in

thedata.

However, trust is not always beneﬁcial. For example,

Padilla et al. (2022a, 2022b) found that viewing lower-

complexity visualizations was associated with increased

trust in the data; however, thisled to poor decision qual

ity. Similarly, McGinnies and Ward (1980) found that

people were more persuaded by a source’s trustworthi

ness than by their expertise, and O’Brien et al. (2021)

found that greater trust in science can make people vul

nerable to believing pseudoscience. It is important to

note that data visualization should not lead to blind trust.

Instead, visualizations should optimize the presentation

of data so that users can critically evaluate information

(i.e., users should “calibrate” their trust: Elhamdadi etal.,

2022a, 2022b; Han & Schulz, 2020). is goal assumes

that the designer wants the user to accurately understand

the data and not misinform (e.g., see Ethical Interaction

eory; Feltz & Cokely, 2024 for further discussion).

Open questions

Our ﬁndings raise several questions for further explo-

ration. First, we did not ﬁnd an eﬀect of condition as it

pertained to real-world decision-making and perceptions

of decisions (i.e., whether one changed their anksgiv

ing or December holiday plans in Study 1). In Study 1, we

ran an exploratory analysisto investigate whether post-

intervention concern mediated the relationship between

condition and real-world outcomes. We found evidence

that,compared to the Static Condition, participants in

the three other intervention conditions (Static + Anec

dote, Narrative Visualization, and Anecdote) reported

increased concern, which was associated with increased

concern about their own anksgiving plans. ese ﬁnd

ings suggest that such interventions may have poten-

tial to inﬂuence real-world decision-making. Future

research should examine the conditions under which

risk communication methods are most likely to inﬂuence

real-world behaviors and decisions.

In Study 1, we found that many participants’ concerns

about COVID did not change from pre- to post-inter

vention. e reason why this was the case is unclear. We

present an exploratory analysis examining some of the

factors that may have contributed to one’s willingness

to increase their concern post-intervention. We found

evidence that participants who identiﬁed with a right-

leaning political ideology were less likely to increase their

concern than left-leaning participants. is suggests that

right- and left-leaning individuals may process data dif

ferently from one another, and that researchers should

consider how to best communicate risk to both groups.

Political partisanship is just one factor that could inﬂu

ence a person’s willingness to update their concern about

a risk in the face of evidence. Future studies should inves

tigate the factors thatinﬂuence one’s openness to updat-

ing their beliefs.

We would like to note that in the speciﬁc example

used in this study, there was an optimal estimation of

risk (i.e., 40%). However, the translation of that esti

mate to concern will vary by individual. For example,

an extremely ris- averse person or someone with a pre-

existing condition may perceive the risk of 40% as greater

than another individualwould. Whether one is overes

timating or underestimating possible risks depends on

many individual factors, as well as the consequences of

misestimating. In the speciﬁc example we present, we do

not think thatparticipants could have been overly con

cerned. People should be very concerned about a 40%

Page 24 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

chance of dining with an individual with COVID or a

hypothetical illness. However, future work could consider

how diﬀerent types of data presentations may inﬂuence

risk perception when the risk is small (such as vaccine

side eﬀects).

In Study 1, we found evidence that although par

ticipants in the Static Condition reported an increase

inperceived risk of contracting COVID, their perceived

concern decreased. is suggests that participants

remembered the risk-related information presented

to them in the intervention (i.e., the risk of contracting

COVID at a anksgiving dinner with 10 people was

40%); however, they were less concerned about the risk. It

is unclear why there was a disconnect between their risk

perception and concern. Future research could use addi

tional methods—like qualitative free-response analysis-

to explain this phenomenon.

Lastly, participants in the Static Condition in Study

1 demonstrated this backﬁre eﬀect, but this was not

observed in the second study. ere were many impor

tant diﬀerences between Studies 1 and 2, namely, Study

1 was conducted in the context of understanding real-

world data, while Study 2 was conducted in the context

of reasoning about hypothetical data. Other notable dif

ferences include the fact that that Study 1 presented data

as a bar chart, while Study 2 presented it as an icon array,

and thatthe Study 1 data analysis only included partici

pants with low-risk perception (i.e., they planned to dine

with others during the 2020 anksgiving holiday). It is

worth investigating the conditions under which these

backﬁre eﬀects occur. We would alsolike to that in Fig.6,

there is an outlier data point for the Static Condition

(decrease in concern by 100 points). We still observed the

data backﬁre eﬀect even after excluding this individual’s

data.

Limitations

While our research showed an advantage for narra-

tive visualizations compared to static visualizations, it is

unclear whether this advantage would also be found in

other scenarios and contexts. e context used here was

one for which risk was (essentially) cumulative. Cumula

tive risk can be especially challenging to comprehend, in

which case, a visualization like the narrative visualization

may have been particularly helpful. It is an open research

question as to whether narrative visualizations would

helpillustrate risks of various magnitudes. Narrative vis

ualizations could also be explored as a tool tohelp people

understand both absolute and relative risk (such as a dif

ference between 2% and 4%, which is a relative diﬀerence

of 100%).

One limitation of our work was our method of meas

uring trust. Participants were given two Likert-scale

items to measure trust in the visualization, measuring

the perceived intentions of the visualization designer and

the perceived accuracy of the data (I feel like the data

are intended to accurately portray the risks of the new

disease/I feel like the data do accurately portray the risks

of the new disease). While it is not uncommon to use

such Likert-scale items in research on trust in research

(for a review, see Elhamadi etal., 2023), other researchers

argue for more holistic approaches that accurately cap

ture the various dimensions of trust. For example, trust

may be assessed with various methods including quan

titative surveys from social sciences, trust games from

behavioral economics, measuring belief updating, and

measuring trust through perceptual methods (Elham

dadi etal., 2022b). Elhamdadi etal. (2023) presented an

integrated framework for trust in visualization, oﬀering

suggestions on how researchers can best capture the vari

ous dimensions of trust. e authors suggest that trust in

visualization is contingent upon cognitive and aﬀective

factors including visualization clarity, usability, accuracy,

aesthetics, and the extent to which the visualization is an

accurate representation of the underlying data (benevo

lence/ethics). While our measuresassessed two facets of

trust, future research in this area should use a more com

prehensive assessment of trust to better understand how

narrative visualizations impact the diﬀerent components

of trust. While not ideal, we do ﬁnd some evidence for

reliability and validity of the measures we present, as the

two items had high reliability (

= 0.91), and our measure

of trust related to understanding, visualization presence,

and concern in the directions one would expect.

We found that trust in the visualization, via under

standing, mediated the eﬀect of the narrative visuali-

zation on posttest concern. However, we found partial

mediation and not full mediation. is raises the open

question of which other aspects of the narrative visuali

zation contributed to its eﬀectiveness. Answering this

question would be useful for establishing design princi

ples for narrative visualizations. For example, it could

be useful to run a version of the study where partici

pants read about how the 40% risk was calculated,but

they do not view the accompanying visualizations.

It is also possible that time on task inﬂuenced subse

quent concern and risk perception. e current study

did not collect data on timin;, however, the amount of

time interacting with each intervention may havehad

an inﬂuence. Lastly, it is unclear if viewing the narra

tive visualization actually improved understanding of

the data-generating process. Future research should

include objective measures of understanding to deter

mine whether actual understanding maps onto per-

ceived understanding.

Page 25 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

Conclusion

e presented work investigated the promise of narra-

tive visualizations for communicating data about cumu-

lative risk. e narrative visualizations implemented in

our intervention included icon arrays and accompany

ing text describing how one’s risk of contracting a dis-

ease accumulates as the number of people one interacts

with increases. Our results provide promising evidence

that narrative visualizations are an eﬀective tool form

itigatingbackﬁre eﬀects by increasing understanding and

subsequent trust in data. is work suggests that misin

formation correction does not need to rely on emotion-

laden content or interventions that aﬀect risk perception

through heuristic thinking; indeed, it appears possible

that narrative visualizations may be an eﬀective way to

improve risk perception using quantitative data.

Supplementary Information

The online version contains supplementary material available at https:// doi.

org/ 10. 1186/ s41235- 025- 00613-w.

Supplementary Material 1

Acknowledgements

The authors would like to thank the National Science Foundation for funding

this research.

Significance Statement

Many people fail to understand that engaging in risky behaviors over time

increases the likelihood of experiencing such risk—termed cumulative risk.

Understanding accumulated risk is crucial for making informed real-life

decisions. For example, in the case of COVID, the likelihood of transmission

from interacting with a single person is small, but interacting with multiple

people increases the likelihood that one of them will have the disease and

transmit it to you. In this study, we examined how to best communicate accu-

mulated risk data with the public, with the goal of increasing concern about

risk as well as understanding and trust in data. We demonstrated that viewing

narrative visualizations—or data visualizations showing how risk accumulates

over time—increased concern about risk by increasing understanding and

trust in the data. This suggests that narrative visualizations should be used

when communicating information about cumulative risk.

Author contributions

M.F. contributed to manuscript preparation and data analysis. L.W. contrib-

uted to experimental design, stimulus creation, data collection, data analysis,

and writing and editing the manuscript. C.H. contributed to experimental

design, data analysis, and writing and editing the manuscript, H.S. and J.K.W.

contributed to experiment design, data analysis, and writing and editing the

manuscript. A.B. and P.S. contributed to experimental design and writing and

editing the manuscript. All authors read and approved the ﬁnal manuscript.

Funding

This work was supported by the National Science Foundation Award

#2030059. (RAPID: COVID-19 Information Visualizations) awarded to P.S. and

J.K.W.

Availability of data and materials

The datasets supporting the conclusions of this article are available in the OSF

repository, https:// osf. io/ k43ev/? view_ only= 67cc6 401b0 69468 89ac4 99fc1

9afec 2f.

Declarations

Ethics approval and consent to participate

All procedures were determined to be exempt and not regulated by the

University of Michigan Institutional Review Board.

Consent for publication

Consent to publish was obtained from all participants.

Competing interests

The authors declare that they have no competing interests.

Received: 9 May 2024 Accepted: 11 January 2025

References

Adams, W., Armstrong, Z., & Galovich, C. (2015). Can students learn from PhET

sims at home, alone? Physics Education Research Conference, College

Park, MD. https:// www. compa dre. org/ Repos itory/ docum ent/ Serve File.

cfm? ID= 13828 & DocID= 4246

Bach, B., Stefaner, M., Boy, J., Drucker, S., Bartram, L., Wood, J., et al. (2018). Narra-

tive design patterns for data-driven storytelling. In Data-driven storytelling

(pp. 107–133). AK Peters/CRC Press.

Betsch, C., Ulshöfer, C., Renkewitz, F., & Betsch, T. (2011). The inﬂuence of narra-

tive v. statistical information on perceiving vaccination risks. Medical Deci-

sion Making, 31(5), 742–753. https:// doi. org/ 10. 1177/ 02729 89X11 400419

Borgo, R., & Edwards, D. J. (2020). The development of visualization psychology

analysis tools to account for trust. arXiv preprint arXiv: 2009. 13200.

Buckley, C., Robles, P., Hernandez, M., & Chien, A. C. (2022). How China Could

Choke Taiwan. The New York Times. https:// www. nytim es. com/ inter

active/ 2022/ 08/ 25/ world/ asia/ china- taiwan- conﬂ ict- block ade. html

Byrd, A., Cai, W., Macdonald, G., Rhyne, E., Throop, N., Ward, J., & White, J. (2022).

The Toss. The New York Times. https:// www. nytim es. com/ inter active/

2022/ 08/ 28/ sports/ tennis/ tennis- serve- ball- toss. html

Cage, F. (2021). Why 5G is slow in the U.S. but will get better eventually. Reu-

ters. https:// graph ics. reute rs. com/ USA- 5G/ jnpwe yldnpw/

Cook, J., Ecker, U., & Lewandowsky, S. (2015). Misinformation and how to cor-

rect it. In R. A. Scott & S. M. Kosslyn (Eds.), Emerging trends in the social and

behavioral sciences. Wiley. https:// doi. org/ 10. 1002/ 97811 18900 772. etrds

0222

Cook, J., & Lewandowsky, S. (2016). Rational irrationality: Modeling climate

change belief polarization using Bayesian networks. Topics in Cognitive

Science, 8(1), 160–179. https:// doi. org/ 10. 1111/ tops. 12186

Cutello, C. A., Hellier, E., Stander, J., & Hanoch, Y. (2020). Evaluating the eﬀective-

ness of a young driver-education intervention: Learn2Live. Transportation

Research Part F: Traﬃc Psychology and Behaviour, 69, 375–384. https:// doi.

org/ 10. 1016/j. trf. 2020. 02. 009

De La Maza, C., Davis, A., Gonzalez, C., & Azevedo, I. (2019). Understanding

cumulative risk perception from judgments and choices: An application

to ﬂood risks. Risk Analysis, 39(2), 488–504. https:// doi. org/ 10. 1111/ risa.

13206

Doyle, J. K. (1997). Judging cumulative risk 1. Journal of Applied Social Psychol-

ogy, 27(6), 500–524. https:// doi. org/ 10. 1111/j. 1559- 1816. 1997. tb006 44.x

Dutta, P. K., Carvalho, R., Ovaska, M., Tai, C., Tennant, S., Weber, M., & Basile, S.

(2019). Reading the Brexit tea leaves. https:// graph ics. reute rs. com/ BRITA

IN- EU/ 01009 2G834N/ index. html

Elhamdadi, H., Padilla, L., & Xiong, C. (2022a). Using processing ﬂuency as a

metric of trust in scatterplot visualizations. arXiv preprint arXiv: 2209.

14340.

Elhamdadi, H., Gaba, A., Kim, Y. S., & Xiong, C. (2022b). How do we measure

trust in visual data communication?. In 2022 IEEE evaluation and beyond-

methodological approaches for visualization (BELIV) (pp. 85–92). IEEE.

Elhamdadi, H., Stefkovics, A., Beyer, J., Moerth, E., Pﬁster, H., Bearﬁeld, C. X., &

Nobre, C. (2023). Vistrust: A multidimensional framework and empirical

study of trust in data visualizations. IEEE Transactions on Visualization and

Computer Graphics.

Page 26 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

Fagerlin, A., Wang, C., & Ubel, P. A. (2005). Reducing the inﬂuence of anecdotal

reasoning on people’s health care decisions: Is a picture worth a thou-

sand statistics? Medical Decision Making, 25(4), 398–405. https:// doi. org/

10. 1177/ 02729 89X05 278931

Fansher, M., Adkins, T. J., Lalwani, P., Boduroglu, A., Carlson, M., Quirk, M., Lewis,

R. L., Shah, P., Zhang, H., & Jonides, J. (2022b). Icon arrays reduce concern

over COVID-19 vaccine side eﬀects: A randomized control study. Cogni-

tive Research: Principles and Implications, 7(1), 38. https:// doi. org/ 10. 1186/

s41235- 022- 00387-5

Fansher, M., Adkins, T. J., Lewis, R. L., Boduroglu, A., Lalwani, P., Quirk, M., Shah,

P., & Jonides, J. (2022a). How well do ordinary Americans forecast the

growth of COVID-19? Memory & Cognition, 50(7), 1363–1380. https:// doi.

org/ 10. 3758/ s13421- 022- 01288-0

Feltz, A., & Cokely, E. T. (2024). Ethical interaction theory. In Diversity and disa-

greement: From fundamental biases to ethical interactions (pp. 211–246).

Springer Nature Switzerland.

Franconeri, S. L., Padilla, L. M., Shah, P., Zacks, J. M., & Hullman, J. (2021). The

science of visual data communication: What works. Psychological Science

in the Public Interest, 22(3), 110–161. https:// doi. org/ 10. 1177/ 15291 00621

10519 56

Freling, T. H., Yang, Z., Saini, R., Itani, O. S., & Abualsamh, R. R. (2020). When

poignant stories outweigh cold hard facts: A meta-analysis of the anec-

dotal bias. Organizational Behavior and Human Decision Processes, 160,

51–67. https:// doi. org/ 10. 1016/j. obhdp. 2020. 01. 006

Galesic, M., & Garcia-Retamero, R. (2011). Graph literacy: A cross-cultural com-

parison. Medical Decision Making, 31(3), 444–457. https:// doi. org/ 10. 1177/

02729 89x10 373805

Garcia-Retamero, R., & Cokely, E. T. (2014). Using visual aids to help people with

low numeracy make better decisions. In Numerical reasoning in judgments

and decision making about health (pp. 153–174). Cambridge University

Press. https:// doi. org/ 10. 1017/ CBO97 81139 644358. 008

Garcia-Retamero, R., & Cokely, E. T. (2017). Designing visual aids that promote

risk literacy: A systematic review of health research and evidence-based

design heuristics. Human Factors, 59(4), 582–627. https:// doi. org/ 10. 1177/

00187 20817 690634

Garcia-Retamero, R., Okan, Y., & Cokely, E. T. (2012). Using visual aids to improve

communication of risks about health: A review. The Scientiﬁc World Jour-

nal, 2012(1), 562637. https:// doi. org/ 10. 1100/ 2012/ 562637

Grice, J. W., Medellin, E., Jones, I., Horvath, S., McDaniel, H., O’lansen, C., & Baker,

M. (2020). Persons as eﬀect sizes. Advances in Methods and Practices in

Psychological Science, 3(4), 443–455. https:// doi. org/ 10. 1177/ 25152 45920

922982

Gustafson, A., & Rice, R. E. (2020). A review of the eﬀects of uncertainty in

public science communication. Public Understanding of Science, 29(6),

614–633. https:// doi. org/ 10. 1177/ 09636 62520 942122

Han, W., & Schulz, H. J. (2020, October). Beyond trust building—Calibrating

trust in visual analytics. In 2020 IEEE workshop on trust and expertise in

visual analytics (TREX) (pp. 9–15). IEEE. https:// doi. org/ 10. 1109/ TREX5 1495.

2020. 00006

Hayes, A. F. (2017). Introduction to mediation, moderation, and conditional pro-

cess analysis: A regression-based approach. Guilford publications.

Hegarty, M. (2004). Dynamic visualizations and learning: Getting to the diﬃcult

questions. Learning and Instruction, 14(3), 343–351. https:// doi. org/ 10.

1016/j. learn instr uc. 2004. 06. 007

Herring, J., VanDyke, M. S., Cummins, R. G., & Melton, F. (2017). Communicat-

ing local climate risks online through an interactive data visualization.

Environmental Communication, 11(1), 90–105. https:// doi. org/ 10. 1080/

17524 032. 2016. 11769 46

Hullman, J. (2019). Why authors don’t visualize uncertainty. IEEE Transactions on

Visualization and Computer Graphics, 26(1), 130–139. https:// doi. org/ 10.

1109/ TVCG. 2019. 29342 87

Hullman, J., & Diakopoulos, N. (2011). Visualization rhetoric: Framing eﬀects

in narrative visualization. IEEE Transactions on Visualization and Computer

Graphics, 17(12), 2231–2240. https:// doi. org/ 10. 1109/ TVCG. 2011. 255

Hullman, J., Drucker, S., Riche, N. H., Lee, B., Fisher, D., & Adar, E. (2013). A deeper

understanding of sequence in narrative visualization. IEEE Transactions on

Visualization and Computer Graphics, 19(12), 2406–2415. https:// doi. org/

10. 1109/ TVCG. 2013. 119

Janssen, E., van Osch, L., de Vries, H., & Lechner, L. (2013). The inﬂuence of

narrative risk communication on feelings of cancer risk. British Journal of

Health Psychology, 18(2), 407–419. https:// doi. org/ 10. 1111/j. 2044- 8287.

2012. 02098.x

Kelton, K., Fleischmann, K. R., & Wallace, W. A. (2008). Trust in digital informa-

tion. Journal of the American Society for Information Science and Technol-

ogy, 59(3), 363–374. https:// doi. org/ 10. 1002/ asi. 20722

Kerr, J., van der Bles, A. M., Dryhurst, S., Schneider, C. R., Chopurian, V., Freeman,

A. L., & Van Der Linden, S. (2023). The eﬀects of communicating uncer-

tainty around statistics, on public trust. Royal Society Open Science, 10(11),

230604. https:// doi. org/ 10. 1098/ rsos. 230604

Kim, Y. S., Reinecke, K., & Hullman, J. (2017). Explaining the gap: Visualizing one’s

predictions improves recall and comprehension of data. In Proceedings

of the 2017 CHI conference on human factors in computing systems (pp.

1375–1386). https:// doi. org/ 10. 1145/ 30254 53. 30255 92

Koerth, M., & Elena, M. (2020). Why even a small thanksgiving is dangerous.

FiveThirtyEight.

Kong, H. K., Liu, Z., & Karahalios, K. (2019). Trust and recall of information across

varying degrees of title-visualization misalignment. In Proceedings of the

2019 CHI conference on human factors in computing systems (pp. 1–13).

Lammers, J., Crusius, J., & Gast, A. (2020). Correcting misperceptions of

exponential coronavirus growth increases support for social distancing.

Proceedings of the National Academy of Sciences, 117(28), 16264–16266.

https:// doi. org/ 10. 1073/ pnas. 20060 48117

Lee, C., Yang, T., Inchoco, G. D., Jones, G. M., & Satyanarayan, A. (2021). Viral

visualizations: How coronavirus skeptics use orthodox data practices to

promote unorthodox science online. In Proceedings of the 2021 CHI confer-

ence on human factors in computing systems (pp. 1–18). https:// doi. org/ 10.

1145/ 34117 64. 34452 11

Levine, A. J., Hernandez, M., & Spring, J. (2021). Vanishing tropical rainforests.

https:// graph ics. reute rs. com/ GLOBAL- DEFOR ESTAT ION/ RAINF OREST/

klvyk zrbxvg/

Lewandowsky, S., Ecker, U. K., Seifert, C. M., Schwarz, N., & Cook, J. (2012). Misin-

formation and its correction: Continued inﬂuence and successful debias-

ing. Psychological Science in the Public Interest, 13(3), 106–131. https:// doi.

org/ 10. 1177/ 15291 00612 451018

Lewicki, R. J., Tomlinson, E. C., & Gillespie, N. (2006). Models of interpersonal

trust development: Theoretical approaches, empirical evidence, and

future directions. Journal of Management, 32(6), 991–1022. https:// doi.

org/ 10. 1177/ 01492 06306 294405

Liew, T. W., Tan, S. M., & Seydali, R. (2014). The eﬀects of learners’ diﬀerences on

variable manipulation behaviors in simulation-based learning. Journal of

Educational Technology Systems, 43(1), 13–34. https:// doi. org/ 10. 2190/ ET.

43.1.c

Lipkus, I. M., & Hollands, J. G. (1999). The visual communication of risk. JNCI

Monographs, 1999(25), 149–163. https:// doi. org/ 10. 1093/ oxfor djour nals.

jncim onogr aphs. a0241 91

Lord, C. G., Ross, L., & Lepper, M. R. (1979). Biased assimilation and attitude

polarization: The eﬀects of prior theories on subsequently considered evi-

dence. Journal of Personality and Social Psychology, 37(11), 2098. https://

doi. org/ 10. 1037/ 0022- 3514. 37. 11. 2098

Magana, A. J., & Silva Coutinho, G. (2017). Modeling and simulation practices

for a computational thinking-enabled engineering workforce. Computer

Applications in Engineering Education, 25(1), 62–78. https:// doi. org/ 10.

1002/ cae. 21779

Mayer, R. E. (2005). Cognitive theory of multimedia learning. The Cambridge

Handbook of Multimedia Learning, 41, 31–48.

Mayr, E., Hynek, N., Salisu, S., & Windhager, F. (2019). Trust in Information Visuali-

zation. In TrustVis@ EuroVis (pp. 25–29).

McGinnies, E., & Ward, C. D. (1980). Better liked than right: Trustworthiness and

expertise as factors in credibility. Personality and Social Psychology Bulletin,

6(3), 467–472. https:// doi. org/ 10. 1177/ 01461 67280 63023

McNaughton, C. D., Cavanaugh, K. L., Kripalani, S., Rothman, R. L., & Wallston, K.

A. (2015). Validation of a short, 3-item version of the subjective numeracy

scale. Medical Decision Making, 35(8), 932–936. https:// doi. org/ 10. 1177/

02729 89X15 581800

Newell, R., Dale, A., & Winters, C. (2016). A picture is worth a thousand data

points: Exploring visualizations as tools for connecting the public to

climate change research. Cogent Social Sciences, 2(1), 1201885. https:// doi.

org/ 10. 1080/ 23311 886. 2016. 12018 85

Nyhan, B., & Reiﬂer, J. (2015). Displacing misinformation about events: An

experimental test of causal corrections. Journal of Experimental Political

Page 27 of 27

Fansheretal. Cognitive Research: Principles and Implications (2025) 10:7

Science, 2(1), 81–93. https:// doi. org/ 10. 1017/ XPS. 2014. 22[Opens inane

wwind ow]

O’Brien, T. C., Palmer, R., & Albarracin, D. (2021). Misplaced trust: When trust in

science fosters belief in pseudoscience and the beneﬁts of critical evalua-

tion. Journal of Experimental Social Psychology, 96, 104184. https:// doi. org/

10. 1016/j. jesp. 2021. 104184

Okan, Y., Garcia-Retamero, R., Cokely, E. T., & Maldonado, A. (2015). Improving

risk understanding across ability levels: Encouraging active processing

with dynamic icon arrays. Journal of Experimental Psychology: Applied,

21(2), 178. https:// doi. org/ 10. 1037/ xap00 00045

O’neil, C. (2017). Weapons of math destruction: How big data increases inequality

and threatens democracy. Crown.

Padilla, L., Fygenson, R., Castro, S. C., & Bertini, E. (2022b). Multiple forecast visu-

alizations (mfvs): Trade-oﬀs in trust and performance in multiple covid-19

forecast visualizations. IEEE Transactions on Visualization and Computer

Graphics, 29(1), 12–22.

Padilla, L., Kay, M., & Hullman, J. (2022a). Uncertainty visualization. In W. Pie-

gorsch, R. Levine, H. Zhang, & T. Lee (Eds.), Computational statistics in data

science (pp. 405–421). Wiley.

Pandey, S., McKinley, O. G., Crouser, R. J., & Ottley, A. (2023). Do you trust what

you see? Toward a multidimensional measure of trust in visualization.

In 2023 IEEE visualization and visual analytics (VIS) (pp. 26–30). IEEE.

Park, S., & Gil-Garcia, J. R. (2022). Open data innovation: Visualizations and

process redesign as a way to bridge the transparency-accountability

gap. Government Information Quarterly, 39(1), 101456. https:// doi. org/ 10.

1016/j. giq. 2020. 101456

Peck, E. M., Ayuso, S. E., & El-Etr, O. (2019). Data is personal: Attitudes and per-

ceptions of data visualization in rural pennsylvania. In Proceedings of the

2019 CHI conference on human factors in computing systems (pp. 1–12).

Petrova, D., Garcia-Retamero, R., & Cokely, E. T. (2015). Understanding the harms

and beneﬁts of cancer screening: A model of factors that shape informed

decision making. Medical Decision Making, 35(7), 847–858. https:// doi. org/

10. 1177/ 02729 89x15 587676

Rakow, T., Heard, C. L., & Newell, B. R. (2015). Meeting three challenges in risk

communication: Phenomena, numbers, and emotions. Policy Insights

from the Behavioral and Brain Sciences, 2(1), 147–156. https:// doi. org/ 10.

1177/ 23727 32215 601442

Reinholtz, N., Maglio, S. J., & Spiller, S. A. (2021). Stocks, ﬂows, and risk response

to pandemic data. Journal of Experimental Psychology: Applied, 27(4), 657.

https:// doi. org/ 10. 1037/ xap00 00395

Rhodes, R. E., Rodriguez, F., & Shah, P. (2014). Explaining the alluring inﬂuence

of neuroscience information on scientiﬁc reasoning. Journal of Experimen-

tal Psychology: Learning, Memory, and Cognition, 40(5), 1432. https:// doi.

org/ 10. 1037/ a0036 844

Rodriguez, F., Rhodes, R. E., Miller, K. F., & Shah, P. (2016). Examining the inﬂu-

ence of anecdotal stories and the interplay of individual diﬀerences on

reasoning. Thinking & Reasoning, 22(3), 274–296. https:// doi. org/ 10. 1080/

13546 783. 2016. 11395 06

Roozenbeek, J., Schneider, C. R., Dryhurst, S., Kerr, J., Freeman, A. L., Recchia, G.,

Van Der Bles, A. M., & Van Der Linden, S. (2020). Susceptibility to misin-

formation about COVID-19 around the world. Royal Society Open Science,

7(10), 201199.

Rutjens, B. T., Heine, S. J., Sutton, R. M., & van Harreveld, F. (2018). Attitudes

towards science. In Advances in experimental social psychology (Vol. 57, pp.

125–165). Academic Press.

Segel, E., & Heer, J. (2010). Narrative visualization: Telling stories with data. IEEE

Transactions on Visualization and Computer Graphics, 16(6), 1139–1148.

https:// doi. org/ 10. 1109/ TVCG. 2010. 179

Seifert, C. M. (2002). The continued inﬂuence of misinformation in memory:

What makes a correction eﬀective?. In Psychology of learning and motiva-

tion (Vol. 41, pp. 265–292). Academic Press. https:// doi. org/ 10. 1016/

S0079- 7421(02) 80009-3

Shah, P., Michal, A., Ibrahim, A., Rhodes, R., & Rodriguez, F. (2017). What makes

everyday scientiﬁc reasoning so challenging?. In Psychology of learning

and motivation (Vol. 66, pp. 251–299). Academic Press. https:// doi. org/ 10.

1016/ bs. plm. 2016. 11. 006

Sloman, S. A. (1994). When explanations compete: The role of explanatory

coherence on judgements of likelihood. Cognition, 52(1), 1–21. https://

doi. org/ 10. 1016/ 0010- 0277(94) 90002-7

Slovic, P. (2000). What does it mean to know a cumulative risk? Adolescents’

perceptions of short-term and long-term consequences of smoking.

Journal of Behavioral Decision Making, 13(2), 259–266. https:// doi. org/

10. 1002/ (SICI) 1099- 0771(200004/ 06) 13:2% 3c259:: AID- BDM336% 3e3.0.

CO;2-6

Slovic, P., Fischhoﬀ, B., & Lichtenstein, S. (1978). Accident probabilities and seat

belt usage: A psychological perspective. Accident Analysis & Prevention,

10(4), 281–285. https:// doi. org/ 10. 1016/ 0001- 4575(78) 90030-1

Slovic, P., Väställ, D., Erlandsson, A., & Gregory, R. (2017). Iconic photographs

and the ebb and ﬂow of empathic response to humanitarian disasters.

Proceedings of the National Academy of Sciences, 114(4), 640–644. https://

doi. org/ 10. 1073/ pnas. 16139 77114

Sterman, J. D. (2011). Communicating climate change risks in a skepti-

cal world. Climatic Change, 108, 811–826. https:// doi. org/ 10. 1007/

s10584- 011- 0189-3

Swire-Thompson, B., DeGutis, J., & Lazer, D. (2020). Searching for the backﬁre

eﬀect: Measurement and design considerations. Journal of Applied

Research in Memory and Cognition, 9(3), 286–299. https:// doi. org/ 10.

1016/j. jarmac. 2020. 06. 006

Thagard, P. (2007). Coherence, truth, and the development of scientiﬁc knowl-

edge. Philosophy of Science, 74(1), 28–47. https:// doi. org/ 10. 1086/ 520941

Tiede, K. E., Bjälkebring, P., & Peters, E. (2022). Numeracy, numeric attention,

and number use in judgment and choice. Journal of Behavioral Decision

Making, 35(3), e2264. https:// doi. org/ 10. 1002/ bdm. 2264

Van Der Bles, A. M., van der Linden, S., Freeman, A. L., & Spiegelhalter, D. J.

(2020). The eﬀects of communicating uncertainty on public trust in facts

and numbers. Proceedings of the National Academy of Sciences, 117(14),

7672–7683.

Van der Pligt, J. (1996). Risk perception and self-protective behavior. European

Psychologist, 1(1), 34–43. https:// doi. org/ 10. 1027/ 1016- 9040.1. 1. 34

Wang, H. C., & Doong, H. S. (2014). Eﬀects of online advertising strategy on

attitude towards healthcare service. In 2014 47th Hawaii international

conference on system sciences (pp. 2725–2732). IEEE.

Witt, J., Hao, C., & Shah, P. (2022). The impact of visualizing the process of

disease spread on social distancing intentions and attitudes. In Proceed-

ings of the human factors and ergonomics society annual meeting (Vol. 66,

No. 1, pp. 2026–2030). SAGE Publications. https:// doi. org/ 10. 1177/ 10711

81322 661172

Yang, F., Cai, M., Mortenson, C., Fakhari, H., Lokmanoglu, A. D., Hullman, J.,

Franconeri, S., Diakopoulos, N., Nisbet, E.C., & Kay, M. (2023). Swaying the

public? Impacts of election forecast visualizations on emotion, trust, and

intention in the 2022 US midterms. IEEE Transactions on Visualization and

Computer Graphics.

Zebregs, S., van den Putte, B., Neijens, P., & de Graaf, A. (2015). The diﬀerential

impact of statistical and narrative evidence on beliefs, attitude, and inten-

tion: A meta-analysis. Health Communication, 30(3), 282–289. https:// doi.

org/ 10. 1080/ 10410 236. 2013. 842528

Zhao, L. (2017). Interface factors that aﬀect users’ trust towards information

visualization (Master’s thesis, Purdue University).

Zipkin, D. A., Umscheid, C. A., Keating, N. L., Allen, E., Aung, K., Beyth, R.,

Kaatz, S., Mann, D. M., Sussman, J. B., Korenstein, D., & Schardt, C. (2014).

Evidence-based risk communication: A systematic review. Annals of

Internal Medicine, 161(4), 270–280.

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in pub-

lished maps and institutional aﬃliations.